The National Highway Traffic Safety Administration (NHTSA) this week warned Autonomous Vehicle (AV) companies that their vehicles’ interference with emergency response operations is “unacceptable.” The federal regulator called on AV system designers and operators to “immediately focus their resources on fixing this issue.”

This got me thinking.

How often do robotaxis interfere with emergency responders? How much do these incidents cost the local community? Who bears the burden?

To earn public trust, AV companies are now being asked to go beyond basic technology (measured in crash rates) and demonstrate tangible social and economic responsibility.

Emergency responders

Let’s start with a dive into robotaxi-emergency responder mishaps.

In the NHTSA letter, administrator Jonathan Morrison flagged as “a disturbing trend … a clear pattern of driverless AVs interfering with law enforcement and other first responders.” He noted multiple instances in which “AVs drove directly into active emergency scenes, blocked the paths of ambulances and firefighters, or failed to recognize and respond to basic safety conditions like flashing lights, flares, smoke, fire, and traffic cones.”

Over the years, AV industry observers—not to mention innocent bystanders—have seen numerous images or videos of robotaxis driving obliviously into trouble. These scenes, however, are anecdotal and sporadic. Often, social media treats them as a source of entertainment, without serious discussion and scant explanation from AV companies. There has been little follow-up by the federal safety regulator— until now.

Public officials are starting to press questions. How often does this stuff happen? How bad are these cases? Why do they keep happening?

I contacted Rachel Castignoli, senior consultant for transportation and public works in Austin, Texas. She’s been spearheading a project to keep tabs on all incidents involving autonomous vehicles that are reported by emergency responders, residents and city employees.

Details are gathered, categorized and fed into a public dashboard here: https://austin.maps.arcgis.com/apps/dashboards/dbd59d9b20634de1afad32a8267b739b#

I filtered the list down to just reports by the Austin Fire Department (AFD), Police Department (APD), Emergency Medical Services (EMS), and Independent School District (AISD) Police, highlighting only the incidents affecting first responders.

From April to June 2026, the city of Austin counted 47 incidents of robotaxi interference with emergency responders. Last month alone (June 2026), the AFD, APD, EMS and AISD police recorded 16 AV incidents.

Castignoli noted that the recorded incidents on the list do “NOT” include those reported only on social media, or self-reported by the AV company.

It’s important to note that it was five years ago when the first robotaxis commercially rolled out on public streets in the United States. Given that experience, the big surprise is that AVs still haven’t learned how to cope with ambulances and fire engines.

As the list documented in Austin shows, mishaps caused by socially awkward robotaxis reported by emergency responders are hardly once-in-a-blue-moon.

Nonetheless, these mishaps tend to be brushed off as occasional, isolated “edge cases.”

NHTSA’s administrator, in his letter, sounded an alarm against this laissez-faire tendency: “Let me be clear: the inability to detect and appropriately respond to such situations represents a functional insufficiency. Emergency scenes are not rare or extreme ‘edge cases’.”

Training offered by AV companies

Right after NHTSA warned AV companies about interfering with emergency responders, a social media post from robotaxi industry leader Waymo touted its “First Responder Outreach” program. The company claimed that it has conducted in-person training of more than 35,000 first responders across the United States.

Education is critical. First responders should know how to interact with AVs.

But shouldn’t AVs know how to interact with first responders? Does Waymo have a “program” for that?

Setting aside that issue, Waymo stressed that its in-person training is complemented by online programs. It offers safety education—free of charge—to professionals who take an oath to serve, protect people and save lives.

The company’s safety training programs cover “how Waymo AVs recognize and respond to emergency operations, detailed instructions for approaching and interacting with the vehicle and critical safety and extrication procedures, including how to disconnect the battery.”

The AV company’s free safety-training pitch doesn’t calculate the time, resources and taxpayers’ money a city must spare, sending its emergency responders to AV school. Castignoli declined to comment on such costs.

Although Texas law does not require AV companies to train first responders, Castignoli noted, “Our first responders undergo a lot of continuous training throughout their careers, and we try to offer many AV training opportunities so as not to burden the system too much.”

Typically, she said, “We set up at least six trainings over three days to accommodate the shift system at the fire department—by offering night training for folks who work nights.” The training, she said, takes an hour.

However, this hour multiplies because every AV company’s different vehicle models demand different procedures for emergency responders.

Given the number of AVs on its streets, Castignoli stressed, “We definitely want the opportunity to train on the vehicles.”

Collaboration between AV companies and cities

The city of Austin is offering AV companies extensive collaboration. Examples include “sharing maps and information that could be useful to AV companies (school zones, valet areas, special event closures) … AV companies do receive communication directly from our 911 dispatch system which should allow them to create timely exclusion zones around major emergencies,” she noted.

Not clear, however, are reciprocal data/information exchanges between cities and AV companies. For instance, AV companies haven’t told Austin how many of their vehicles are in operation.

Texas state law requires each AV to be registered with the DMV. “We know how many are in the state, but not how many are operating in Austin,” Castignoli said.

This matters.

“There is data that would help us improve our policies and procedures around AVs and separate data that could help us with transportation planning more broadly,” she said. “If we know the location of AV issues (disengagements, etc.) we can work on the infrastructure or other issues in the road that may have caused it.”

This need motivated the city to develop an incident map, explained Castignoli, hoping to enable AV companies “to identify patterns,” and figure out “how to fix them.”

911 calls

Less publicized by the media is the issue of robotaxi passengers, impaired by drugs or alcohol, who fall asleep or pass out, and don’t exit the vehicle at their destination.

A robotaxi with no human driver cannot diagnose an inert passenger. The cab’s remote assistant, hundreds or thousands of miles away, cannot tell if the passenger is breathing. The robotaxi’s programmed solution is a 911 call. Austin’s first responders handled 99 of these mostly false alarms between March and December of 2025. Each, however, had to be treated as a cardiac arrest dispatch involving multiple ambulances, a fire truck, and sometimes police.

Only two to three percent of these passengers needed to be transported by EMR.

Inconsistent information from AV companies hinders emergency responders’ ability to triage the situation with the right resources.

Such spurious emergencies are not a freshly discovered inconvenience. Austin broached the issues with Waymo more than two years ago. In a Special Joint Meeting of the city’s Public Safety and Mobility Committee last April (https://austintx.new.swagit.com/videos/384367 ), EMR officials noted, “We’re still unclear on the capabilities of their remote support to communicate with the individual in that vehicle, whether that’s through noise communication, whether they have a video feed in the vehicle, or if they can, they can make noises inside of the patient compartmentalize the world.”

Accountability

Castignoli stressed that Austin has “good communication” with AV companies.

But she indicated, “We continue to struggle with not really having a way to hold AVs accountable for issues.”

A recent change to Texas law allows Austin to report unsafe AV behavior to the Texas DMV. “That is great, “she noted. “But it is frustrating to have issues repeat themselves when we have communicated clearly that we need AVs to do better (with hand signals, school buses, etc).”

Individual cases reported in Austin last month

Here are examples of recent incidents in which AVs were cited for interfering with Austin’s emergency responders last month.

• June 30, 2026, AFD report

  At approximately 2305 hours a WAYMO utilized a private parking lot to circumnavigate a police barricade on a fatality collision. It then turned onto the closed section of roadway and started driving towards the scene itself where Officers were on foot and the body was still present. I stopped the vehicle and was able to call the WAYMO hotline to have it turned around…

• June 23, 2026, APD report

  While working an intersection closure for a traffic light, between 9:30 and 9:35am today, at Schriber turning Eastbound onto Oltorf, I was giving the Robotaxi vehicle very clear directions to turn right (Westbound), but it was hellbent on turning only left and seemed very confused at hand signal directions. This was a new black Model Y that had no driver or passenger and was marked with “Robotaxi” emblems on the sides. This was not one of the gold colored prototype Robotaxi vehicles that we’re seeing in more frequency. The vehicle never complied with hand signals and we eventually allowed it to turn.

• June 19, 2026, AFD report

  ENG13 was responding to a motor vehicle collision. While making a left turn to the accident a WAYMO drove forward and clipped the side of the engine with its left rear camera.

• June 13, 2026, AFD report

  ENG01 was attempting to make a right turn from E 8th street to northbound Red River Street. Waymo was headed southbound on Red River Street. ENG01 turn took them partially and briefly into southbound traffic in front of the Waymo. Autonomous vehicle could not figure out how to get out of the way of ENG01 while emergency lights were on. Autonomous vehicle attempted to back up and move forward several times, blocking both directions of traffic until eventually backing up and driving backwards and northbound Red River until it found a spot to get out of the way.

• June 10, 2026, APD report On 6/10

  Vehicle stopped abruptly in intersection; I blocked oncoming traffic; before I could contact Waymo, vehicle continued and almost struck another motorist; I stopped the vehicle and made contact with operator Caleb 3334128. Lt White: “This occurred at E 6th St and San Jacinto. I confirmed that the Waymo was in the intersection. A vehicle was about to turn right on red when the Waymo suddenly shot forward almost striking the vehicle that managed to brake in time.”

• June 2, 2026, AFD report

  Engine 6 was responding code 3 to a traffic injury. We were traveling in the left lane and the WAYMO was in the left lane in front of us. They are usually really good about moving to the right, however, this time, it seemed as if it could not make up its mind. The WAYMO slowed down, turned its hazards on, then turned left slightly, then back to the right, then to the left, and came to a stop. Engine 6 did not pass the WAYMO, we were unsure of its next move, until it finally decided to pull all the way to the right.

Bottom line:

With the NHTSA’s latest letter, AV companies are clearly put on notice. Most likely, recalls are the only pathway left for robotaxi companies to take in responding to NHTSA’s request.

But this comes with some serious concerns and challenges.

If the past is prelude, can recalls actually fix the problems of AVs interfering with emergency responders?

Thus far, certain workarounds and software updates introduced by AV companies to prevent their vehicles from driving into flood water, passing school buses with a flashing red light, or running into construction zones have not worked. Given repeated offences, AV companies can’t declare their software patches successful.

Adding even more complexity for robotaxis are emergency situations unfolding and altering the physical world on a minute-by-minute basis.

Expect a long journey for AVs to become careful and competent drivers.

Brain-inspired computing is a long-established engineering field, going as far back as the 1940s. The idea of emulating the human brain’s architecture, or at least the actions of individual neurons, has attracted neuroscientists and computer scientists seeking radically new ways to look at computing.

R&D activities on biological computing using actual neurons, or neuromorphic computing using hardware or software inspired by neurons, both of which draw inspiration from brains, began well before ChatGPT muscled Artificial Intelligence into the vernacular.

Nevertheless, brain-inspired computers have never become broadly available in the commercial market, nor have they rivaled the performance of software and hardware built by established computing. That may be about to change.

Today, despite its sudden ubiquity, AI is facing a backlash of fear and controversy. Scientists, engineers and the public alike have noticed with alarm AI’s voracious and expanding appetite for compute power. The future of AI may depend not upon further scaling, with even more GPUs and even more power, but upon finding ways to reduce the amount of computing necessary to achieve a quality result.

And one of the best arguments for neuromorphic computing is the billion-fold difference in energy efficiency between brains and datacenters. The human brain has been measured to run on about 20 watts. A massive GPU-infested datacenter processing an LLM mimicking human cognition needs megawatts to gigawatts. This disparity exists in part because of the brain’s slow, non-linear, event-driven spiking, compared to a computer’s frenetically continuous clock-driven operation, in which sometimes every node in a circuit can change on every clock cycle.

Not only is scaling up to greater power causing public outcry, but computer scientists themselves are increasingly skeptical that larger, more complex models will improve the problems with today’s models: issues like hallucinations or poor-quality long videos.

A further roadblock looming across the future of AI is the learning problem. Today AI models are trained at great expense in datacenters. They cannot learn continuously from experience. So when a model encounters a new environment, or even a novel prompt, designers must collect masses of data on the new situation and retrain the model. Not only is this a slow, expensive, and necessarily infrequent process, but there is worse news. As models are retrained, they can increasingly and unpredictably forget things they already knew, producing hallucinations where once they produced useful output.

In contrast, humans learn quickly—often from a single experience and without outside guidance. And as we learn more, the experience enriches, not destroys, our existing knowledge. We don’t forget what a pedestrian looks like when we learn about a detour sign. Nor do we lose track of objects or perspective in long video sequences. These skills are still not well understood.

Brain-derived computing may hold some solutions.

Software launch

Against this backdrop, The Biological Computing Co. (TBC), a four-year old San Francisco-based startup, is seizing the moment with an effort to bring to fruition a biological computing platform. But unlike the trail of failed neuromorphic predecessors, TBC’s first product and its path to revenue, slated for launch this year, is not a piece of hardware. It’s software.

The startup is applying the principles of brain-inspired computing to AI software, such as generative video, computer vision, and AI models. TBC is promising to deliver software tools that, for instance, help advertising agencies or marketing companies create higher quality AI-generated video faster and cheaper. They are promising a difference users will pay for.

During our recent interview, Alexander Ksendzovsky, TBC’s co-founder and CEO, made clear that real-time biological computing remains the company’s core and eventual ambition. TBC does, in fact, build and operate silicon chips with living neuron cells on them. But the chips are not used for computing.

Instead, TBC studies the behavior of the neurons on these chips to derive algorithms that resemble what the neurons are doing. Then they code up these algorithms and insert them as adapters into existing AI models at key points. In effect, these adapters modify the data flowing through the AI system’s machine-learning network to make the AI system’s tasks of inference and memory much easier.

That can improve both efficiency and quality of results. Users can plug in these neuron-derived adapters and reap the benefits in speed, power consumption, or quality of results, without redesigning or retraining their original model or altering the way they use the AI tools.

Who’s on the team?

Neither TBC’s co-founders, Ksendzovsky. nor COO Jonathan Pomeraniec is a computer scientist. They are neurosurgeons and neuroscientists. They still do brain surgery, but only sporadically these days.

The two co-founders have been tracking the development of biological computing for 20 years. They have teamed with John Wittig, trained as a computer engineer, chip designer, and neuroscientist. Wittig is TBC’s vice president of products and engineering.

Wittig brings to TBC his solid hardware expertise. To earn his PhD, he built microchips that model the brain and wrote software to model how brain cells process information.

Asked about his research focus, Wittig said that he’s been working to understand “the basis of the computation that allows us to interact with the world.”

How does it work?

The startup has acknowledged that its work stands on the shoulders of the accomplishments of many brain computer interface companies. Although the commercial market remains largely devoid of brain-inspired computers yet, the field’s R&D activities have never slowed. Ksendzovsky said. “Neuroscience has already massively flourished, which led to our much better understanding of how neurons process information.”

During the interview, Ksendzovsky showed off TBC’s chip. The chip shown to us did not have brain cells on it, but the company has already built brain cells on such a chip. That working chip has helped the team learn fundamental principles about how the brain processes data.

Today, “we can grow neurons long enough, we have enough electrodes, we have computational tools to be able to understand brain signals,” explained Ksendzovsky. With this arsenal of neurons on the working chip, the company, over the last four years, has gained “incredible access to how brain cells actually compute information.”

The team uses the electrode interface on the chip to stimulate the neurons on the chip’s surface. For example, said Ksendzovsky, “Assume the information is a picture of my face. The pixels correspond to the electrodes, and we can send that pixel information as currents, through the electrodes and into the biological network.”

He went on, “Then we can record the responses of the neurons, thus allowing us to understand and model mathematically how brain cells process an image, or any data that we put into the dish.”

Add this processing into transformer architecture, for example, and it enables an AI model to look a little more biological, noted Ksendzovsky. “It’s not a complete redo of the hardware,” but making the software “slightly more biological … actually leads to incredible improvement in efficiency,” in terms of training and generating video.

Proof point

TBC’s first proof of concept was the performance of video-generation models. Today’s AI-based video generation can create quite impressive short sequences. But longer sequences get increasingly inaccurate and can eventually lose track of objects or become meaningless blurs. Adding their neuro-derived adapter into the AI model, “we generated five to 10 times more video … at a higher quality.” TBC succeeded by simply “making the video generation algorithms a little bit more biological” by incorporating TBC’s adapters and learning rules, noted Ksendzovsky.

Today, TBC’s team is doing targeted experiments to exploit some fundamental principles in biology. “We observe them, mathematically model them, and apply them in a very directed way toward whatever problem we’re trying to solve.”

Claiming an accelerated discovery loop the company has pulled off, the CEO remarked, “In just a short period of time, we’ve gotten incredible progress in terms of what we’re able to do.”

Source: The Biological Computing Co. (TBC)

How to commercialize it

The issue then comes down to how TBC plans to commercialize its software. What’s the business model?

People can access video generation models that have been enhanced and optimized by using TBC’s adapters, “either through the marketplace or through an API that goes to TBC’s website,” explained the CEO.

Typically, TBC starts the optimization process by using a model made available on open source. It can also work on a closed source model provided by a commercial client, explained Wittig. TBC applies the company’s own transformation to the model, which remains closed.

In short, the only access is through APIs. People anywhere in the world, at any time, can access TBC optimized inference models in an online marketplace, paying the marketplace directly per token and per model usage, Wittig explained. “Then we share a portion of that revenue.”

TBC’s COO Pomeraniec stressed, the key advantage of the startup’s product is that it plugs and plays into current AI workflows. “People don’t have to change their behavior” to adopt TBC’s software, even though it was built on brain-inspired computer, a radically different compute paradigm. “The end user doesn’t have to change anything.”

TBC raised a $25 million seed led by Primary Ventures in March.

— Additional reporting by Ron Wilson.

Robotaxis are often judged by whether they are safer than human drivers.

It is a necessary question, but too narrow and too low a baseline.

“Safer than humans” is an incomplete public bar for a technology wrapped in decades of science fiction. People expect robots to be more precise, predictable, courteous, and consistently competent. A robotaxi that outperforms the average distracted, impaired, fatigued, or novice driver may clear one technical benchmark while still falling short of public expectations.

That expectation may be unfair. But it is real.

The more useful question is whether automated vehicles currently on our roads are maturing in ways the public system can measure, validate, and trust.

In other words, who grades the robot driver?

The Baseline Cannot Be the Finish Line

Comparing robotaxis to human drivers matters. But humans are not a single benchmark. A distracted teenager, a professional fleet driver, and an alcohol-impaired driver are not equivalent. Averages can hide the conditions that matter most to communities.

The real test is whether or not we see continuous improvement.

A robot driver is not judged like a human who has a bad day. It is judged like a system that should learn from every deployment, software update, edge case, and near miss. Are hard-braking and near-conflict events declining? Are remote assistance calls becoming less frequent within the same operating domain? Are identified problems being fixed or just becoming less visible?

This is where independent data becomes critical.

Governments Cannot Grade What They Cannot Measure

If robotaxis are going to mature in public, governments need enough high-fidelity data to grade performance and measure progress.

That does not mean cities, states, or federal agencies need access to every proprietary detail inside an automated driving system. But they do need consistent, validated measures of how these systems behave in the real world.

Without tangible evidence, public agencies are left grading the robot driver from sparse data. They see visible incidents, hear from residents, read company safety reports, and react to political pressure when a story breaks. Those inputs matter, but they are not enough to support mature governance.

A serious public system needs measures that are consistent across operators, comparable across locations, and stable enough to show improvement over time. Those are grading questions and trust questions.

Trust does not come from novelty or corporate confidence alone. It comes from consistent performance and independent validation. That is the foundation on which conditional permits should be built.

A permit can begin with limits, be relaxed when evidence supports it, and be tightened when recurring problems appear. But that only works if regulators can see what is changing.

The grade cannot be based on press releases, selective disclosures, or ad hoc observations. It has to be based on measurable, validated performance over time.

A Learner’s Permit Is a Data Problem

The learner’s permit analogy matters because it gives us a better way to think about automated mobility.

We do not treat teen drivers as fully mature the day they get a license. We know exposure matters, and that performance in one setting does not automatically mean readiness in every setting. So we use graduated licensing. We allow new drivers onto public roads with legal limits and parental oversight that relax as maturity is demonstrated. The rules help teens gain experience step by step while making responsibility visible to both young drivers and the adults who supervise them.

Automated mobility should grow in the same way.

A robotaxi service might begin in a defined operating domain, with limits on road type, speed, weather conditions, or construction-zone exposure. As evidence improves, the permit could expand. If performance slips, the permit should narrow.

That is not punishment. It is governance. Teen drivers need supervision as privileges expand. Robotaxis need independent evidence, clear operating limits, and an operator responsible for making practical safety decisions within those limits, including when conditions require more restrictive operation.

Conditional permits are the modern automated-vehicle version of a learner’s permit. Independent data is the ongoing road test that helps determine whether privileges should expand or narrow over time.

Existing Tools May Help

We do not need to start from zero.

Existing advanced fleet-data systems could help regulators, insurers, and communities assess whether robotaxis are ready for broader deployment.

NHTSA’s Standing General Order is a valuable start. It gives federal regulators visibility into reported crashes involving automated driving systems and advanced driver assistance systems. But it was not designed to bear the full burden of public trust for automated mobility. It is centered on reportable crashes, not the broader behavioral signals communities need to decide whether automated vehicles are maturing.

A third-party data layer could help fill that gap.

Companies such as Samsara, Motive, and Nauto already provide tools that combine telematics, video, AI-enabled event detection, and operational analytics. These systems were built primarily for human-driven commercial fleets, not as complete validation tools for robotaxis. But that is exactly why they may be useful.

They offer a practical evidence layer that already exists in the market.

No single platform is likely to be perfect. Sensors may miss events, definitions may differ, remote assistance activity may not be fully visible, and scoring models built around human drivers may not transfer cleanly to automated systems. The data would need careful interpretation.

But independent, high-fidelity visibility into vehicle behavior would be far better than relying only on company safety claims, selective public releases, or crash reports.

Many telematics systems already support first notice of loss and rapid event notification. In principle, a governing body could be added as an authorized notification party for certain safety-relevant events. That would reduce dependence on self-reporting alone. The alignment between a third-party system and an automated vehicle’s internal sensors would not be perfect, but even an imperfect outside signal could help focus attention on the cases most in need of review.

A jurisdiction need not choose a single technology. Data aggregators could allow operators to select among approved vendors, while a common layer normalizes inputs for insurers, regulators, or other authorized users.

This would not replace engineering validation, simulation, closed-course testing, or regulatory oversight. But it could give stakeholders a more grounded evidence base.

Safety Is Bigger Than Avoiding Crashes

Crash reporting is necessary, but it is not the whole picture.

A community should not have to wait for serious crashes to learn that a fleet has trouble under certain conditions. Leading indicators should assess how vehicles behave before something goes wrong.

Those are learner’s permit questions.

A human driver does not graduate simply because they avoided a crash during the first few weeks. Responsible parents look for judgment, maturity, and the ability to handle more complicated environments. Automated systems should face the same evidence test. When risk indicators decline, operations can expand. When warning signs persist, they should not.

Safety should include new forms of value, not just crash avoidance. A vehicle that detects passenger distress and routes toward help is doing more than avoiding a collision.

One event may be noise. A pattern is evidence.

Trust Is the Real Scaling Constraint

The latest JD Power Mobility Confidence Index research, conducted in collaboration with MIT’s Advanced Vehicle Technology Consortium, shows that consumer confidence in self-driving vehicles remains stalled. Awareness may be improving, but trust is not keeping pace.

Negative news travels quickly because people judge these systems not only by aggregate performance, but by whether what they see on the road looks competent and acceptable.

Governance matters because it helps technology survive the inevitable news cycles. It gives communities tools short of prohibition, and lets responsible operators distinguish themselves. Regulators get more than headlines to react to. Consumers get evidence that someone other than the operator is watching the system.

That is why independent data matters.

Not because it solves every technical problem. Not because third-party telematics systems were designed to certify robotaxis. Not because every metric will be perfect.

It matters because trust requires more than transparency. It requires evidence that can be checked by someone other than the company asking to be trusted.

A Road Test That Evolves

The answer is not to stop automated mobility. The answer is to help it mature.

Technical performance and public trust are related, but they are not the same. Safety claims are necessary, but not sufficient. Real-world data matters only if it is broad enough to capture what communities actually experience.

The better path is governed learning.

Let these systems operate. Let them improve. But require real-world evidence and independent evaluation, so communities can determine whether the improvement is real.

The same logic will apply beyond robotaxis. Driver assistance, automated trucking, delivery robots, and every software-defined mobility system will face the same question: how do we know they are improving after deployment?

A learner’s permit creates a structured path for teenage drivers to mature. Automated mobility needs the same: a path where privileges are earned, and someone is keeping score.

None of this requires waiting for federal action or perfect data. The tools largely exist. State and city leaders issuing these permits should start the debate now.

—Guest columnist Bryan Reimer is a research scientist at MIT and co-author of How to Make AI Useful. His work focuses on how artificial intelligence and automation shape real-world systems, and he advises industry and policymakers on their deployment. This piece is co-posted by Reimer on LinkedIn.

The autonomous vehicle (AV) industry loves pitching the line that their computer drivers will never get drowsy, distracted or drunk.

But what about degraded?

Calibration drift, often triggered by “shock and vibration, sensor noise accumulation, temperature fluctuations and general material aging,” can pose degradation issues both in hardware and software, explained Sparsh Gautam, an industry expert in computer vision, industrial automation and robotics.

In our recent interview, Gautam warned that such calibration drifts could become “a silent bottleneck” for an industry eager to accelerate the proliferation of robotic systems but challenged to verify the safety of highly automated vehicles.

Gautam has worked with a variety of vision inspection systems in industrial systems. He noted that some sensor systems that began performing at over 90% efficiency slipped below 10% in a matter of months.

With Physical AI now on the agenda, underestimating calibration drift could bring unintended consequences to highly automated vehicles, robotic systems and automated manufacturing processes.

Sensor drifts in automotive

It’s fair to ask to what extent is calibration drift already affecting automotive?

“At least in the early days, I heard that robotaxis had to be recalibrated daily,” said safety expert Phil Koopman, a professor emeritus at Carnegie Mellon University. “But I don’t know what the deal is these days.”

Koopman said,

“The calibration drift problem is as good or as bad as companies want it to be. Spending a bit more money can often permit self-calibration at the start of every driving cycle or buying a sensor that has less tendency to drift.”

Sensor systems in AVs have “definitely improved compared to the manual process that used to happen,” confirmed Gautam, thanks to “the maturing of robust perception pipelines, and multi-sensor calibration stacks.”

Robert Stead, managing director at Sense Media, concurred. “The stacks [in robotaxis] are much more robust, with more redundancy across sensor modes to significantly improve fleet efficiency and reduce the need for cleaning and calibration.”

Although more robust sensor stacks help robotaxis cope with sensor drifts, it is becoming “increasingly more important to integrate the detection and and monitoring of sensor drifts in the workflows,” noted Gautam.

Meanwhile, Stead said, “Compared to robotaxis, Advanced Driver Assistance Systems (ADAS) consumer vehicles have fewer sensors, and are not managed by a fleet owner in the same way.” They tend to “rely on the vehicle owner to follow the maintenance schedule. And service intervals are only once a year or so.”

Stead cautioned, “If you’re driving a car that’s a few years old or more (or even a few days old), there’s currently no way to know, as the driver, if the ADAS system is still calibrated to the standard that it left the manufacturing plant.”

The system may continue to perform as intended for the typical ten-year lifespan of a consumer vehicle, said Stead. “But it might degrade to 90%, 80% or less, and you’d have no way to know until the system was triggered in an accident.”

The threat of uncalibrated sensors is not just hypothetical.

Benjamin May, CEO at AMX13, a German consultancy, noted, “A study some years ago with Belron (car glass company), showed that time to collision is heavily impacted in ADAS features such as an emergency braking assist.” This happens with a camera (or other sensor) that does not know it is de-calibrated. “More complex ADAS features suffer even worse,” he added.

Sustainability of ADAS vehicles

The lifecycle of sensor systems in ADAS vehicles is oft discussed in the context of post-accident vehicle repairs.

For consumers, the immediate headache after an accident is the high cost of repair.

As more vehicles are equipped with cameras, radars and sometimes even lidars, consumers today could face a repair bill, just to fix a fender bender for their ADAS vehicle, as high as $7,500.

Even less discussed is what could be a serious safety problem after a repair. If not properly recalibrated, an ADAS vehicle—touted by OEMs as a quantum leap in safe driving—could end up with its sensors all looking the wrong way.

Repairing today’s sensor-laden ADAS vehicles no longer requires just parts replacement and body cosmetics. More important is the precise step-by-step measurement and calibration of sensors stipulated by each automaker.

Unfortunately, very few of today’s mom-and-pop auto body shops can handle that.

AMX13’s May believes automakers—typically too busy with what they see as the latest burning issues—haven’t fully paid attention to serious problems of sensor drift, including who is responsible.

May asked, “Is sensor drift a topic that must be dealt with by OEM, aftermarket, or suppliers (tier ones)?”

Why is calibration hard?

Post-crash ADAS repair is hard enough. Even harder is for automakers to acknowledge, and then deal with, sensor calibration drift that happens silently— unbeknownst to anyone.

Stead asked: “How do you compare the vehicle’s ADAS performance today vs how it would have performed fresh from the factory?”

There’s no baseline to enable meaningful comparison, said Stead. “In a way, you just have to hope that all the sensors are still working well which will contribute to your safety in some way.” In short, automakers are asking consumers for blind faith, although they themselves have no idea what’s going on. There are no mechanisms to alert automakers and drivers about calibration drift.

Gautam explained that drift can be generated by both hardware and software.

“Slight drifts in vision system performance from a camera or vision sensor are often telling of a specific failure mode,” Gautam noted. Complications arise “when we suddenly run into very low yield or poor camera performance.” This “indicates compounding failure modes contributing to the low yield.”

Drift contributed by software

In general, drift is triggered over time by shocks and vibration, sensor noise accumulation, big temperature fluctuations and plain old aging.

However, “the software layers also have potential to cause significant drift,” noted Gautam, “owing to the environment it learns from and the targets/reference points it picks up to internally calibrate the camera for a use case.”

More specifically “Failure modes that arise from software include overfitting a calibration procedure to a specific checkerboard or to a specific environment target which can then break with variation induced from lighting/orientation changes,” Gautam added. Another issue is that “in the industry’s efforts to avoid calibration bottleneck, it is becoming increasingly popular to constantly recalibrate on the environment as we go.”

Gautam said, “Although the algorithms and software layer is a more difficult problem to solve because the environment is never static, the more complex root-causing problems actually stem from a mixture of failure modes.”

For example, if lighting changes affect a sensor’s input reading or if it’s covered by dirt, the sensor’s contrast and overall field-of-view can be altered. When this happens, software struggles to pick up on specific pixels for autocalibration. The result is misalignment of the camera’s parameters, which can completely distort its vision, he explained.

Are calibration tools available?

Despite the obvious need for tools to calibrate sensors, the auto industry today does not appear to be fully prepared.

Stead suspects that not many players in the auto industry see recalibration as a big deal, possibly because they are unfamiliar with the issues.

Nonetheless, some tier ones are working on this, he said. There are also “third-party solution providers such as Obsurver.”

Stead believes Obsurver has a patented methodology and software solution for evaluating either on-the-fly or in a workshop environment whether sensors are performing to spec.

He said, “This could be integrated into the vehicle so that the vehicle self-evaluates on an ongoing basis and can alert the driver if the calibration has drifted” too far.

Alternatively, a workshop-based test could be integrated into the regular service at the OEM dealership or included in annual roadworthiness testing, Stead suggested.

Until calibration monitoring is regulated or included in New Car Assessment Program (NCAP) requirements, for example, it’s hard to envision that ADAS vehicles will have effective safeguard solutions adopted across OEM production models.

Technically speaking, however, the industry does not lack solutions to sensor drift.

Agreeing with Koopman’s statement that “the calibration drift problem (in automotive) is as good or as bad as automakers want it to be,” Gautam said, “It’s entirely dependent on how automakers and/or component designers factor in for the variations of the failure modes right from the start.”

Solutions such as “more rigid sensor mounts for vibration, improved designs for thermal management, making sensors with materials better rated for environmental wear are component design and manufacturing constraints on the hardware side to improve performance,” Gautam explained. “Even building better synchronized software architectures and having automated periodic diagnosis checks to monitor calibration health is a big leap.”

Simply “performing calibration” is not enough. Vision systems designers must understand the need “to actually make the system preserve the calibrated stack and setup,” Gautam concluded.

Collaboration with academia, and standardization

In the emerging era of Physical AI, developing robust calibration is a prerequisite to high performance in ADAS, humanoids, warehouse robotics and other automated systems. Gautam advocates closer industry-academia collaboration efforts to help standardize calibration operations.

Gautam, for example, would like to see from an engineering framework, a standardized process for the calibration health metrics, such as localization consistency or confidence in drift assessment.”

Also subject to standardization are objects used for recalibrating on the fly, Gautam added. Each player now has its own targets, checkerboards at factory calibration, road features/lanes, SIM-based testing and so on. “But if there are commonly known features that everyone can benefit from, it simplifies the validation methods used across multiple industries that leverage vision systems,” stressed Gautam.

Bottom line

Automakers seem to be dragging their feet, but the industrial market’s need to mitigate calibration drift is real and urgent.

The awareness of calibration drift is indicated by an uptick on job postings for software engineers with expertise in sensor testing and calibration at leading companies in AV, robotics and manufacturing, such as Waymo, Aurora, Toyota, Tesla, Anduril and Applied Intuition.

Related story:

Robotaxis are forcing a question many cities aren’t equipped to answer: how do you govern vehicles that can operate without a driver but still need public oversight?

The debate, so far, has been mostly binary: yes to robotaxis or no. Cities need a third option.

Automated vehicles need room to evolve. But cities also need the authority to slow, pause, or narrow deployment when safety risks emerge. Cities need tools for dealing with other problems beyond overt safety: vehicles traveling without passengers, curb use, neighborhood staging, and the quiet conversion of public streets into fleet infrastructure.

This governance gap is becoming harder to ignore. New York, Pennsylvania, Washington, D.C., Illinois, and Massachusetts are each facing versions of the same basic question: if highly automated vehicles are allowed on public roads, under what conditions should they operate? Details differ. Massachusetts, including Boston, is moving cautiously. Pennsylvania, through Philadelphia’s debate, is exposing the limits of local control when state law holds the main regulatory authority. New York has allowed limited testing with in-vehicle safety operators but has not yet opened the door to paid driverless service. Washington, D.C., is considering steep upfront costs and staged permitting. Illinois is considering a pilot structure that could make high-population counties and select named regions the first test grounds.

The federal government may eventually set clearer national rules. It should. Federal minimum standards matter. So do data requirements, recall expectations, and basic accountability when things go wrong. The public should not have far less visibility into safety and performance in one state than in another simply because reporting rules differ. But cities and states cannot wait for a perfect federal framework. The vehicles are already arriving, and the operational questions are local. The trust questions are immediate.

A flooded roadway is local. So is a blocked intersection, a school zone, an emergency response, or a residential street that suddenly becomes a waiting room for empty vehicles.

That is why mayors and governors need more than slogans about innovation. They need tools they can actually use when technology meets the street.

Earned Scale

The right model is neither prohibition nor a blank check. It’s a conditional permit. It would let an operator begin under defined conditions, expand when the evidence supports expansion, and face temporary restrictions when safety risks arise. When the evidence shows the system is stumbling, pump the brakes until the problem is understood.

That isn’t anti-innovation. It’s how complex public systems learn, and how they build trust before trust is lost.

When safety risks justify it, a city should be able to require a corrective-action plan, narrow the operating domain, reduce fleet size, pause service in certain conditions, suspend expansion, or order a temporary standdown. If the problem is heavy rain, construction zones, flooded roads, emergency scenes or a particular type of intersection, the response should be specifically tailored to the risk.

One approach in Washington, D.C., shows both the promise and the risk in this moment. Its proposed framework, B26-0684, points in the right direction by recognizing that robotaxi governance needs more than a one-time rubber stamp. The bill would create staged permitting, require operational reporting, initially cap commercial AV fleets, and give officials authority to restrict, suspend, or revoke permits when operations create unreasonable public safety risk. That is the right instinct.

But the structure also shows the danger of confusing governance with a toll booth. A $1 million application fee and a $5 million permit fee may keep smaller firms out without necessarily making the system safer or more cost-effective for riders. Meanwhile, a 15-cent-per-mile-traveled fee may not fully reflect the ongoing public costs of road use, curb use, congestion, space for empty vehicles in the right-of-way, and city oversight. A better model would make entry reasonably accessible, then increase obligations as operators scale, create risk, use public space, or impose costs on the city.

Illinois shows a different version of the same tension. The proposed Autonomous Vehicle Pilot Project Act, SB3392, would limit early pilot deployments largely to counties with at least 1 million residents, as well as several specifically named counties. That may make commercial and political sense. But it also reveals the tradeoff at the center of this debate. If safety learning were the only consideration, one might expect pilots to start in smaller, less congested places and work up to dense urban environments, not the other way around.

A vehicle that cannot reliably handle a flooded roadway is a safety problem. A fleet of empty vehicles circling a neighborhood while waiting for rides is a public-space problem. Both matter, but they require different solutions.

Evidence Before Trust

Waymo’s recent flooding-related recall shows why this authority matters. In a voluntary recall filed with federal safety regulators, Waymo reported that software in 3,791 automated driving systems could allow vehicles to slow and then drive into standing water on higher-speed roadways. Waymo also temporarily paused some operations while making safety updates. Those are responsible steps. But they also raise the question: should the public have to rely on the operator alone to decide when a risk is serious, when operations should pause, and when the problem has been fixed?

Cities and states need enough supporting evidence to answer basic questions: Did the patch work? Did the operating domain change? Could the same risk reappear under similar conditions? A software update may fix a technical defect. It does not automatically solve the public trust problem.

Relying on a “Trust me” assurance from corporations is not governance. It asks cities, states, and federal officials to yield public responsibility to the very operators they are supposed to oversee. That may be convenient for deployment. It is not a durable foundation for public trust.

The Street Is Part of the System

Empty vehicles circling, waiting, or concentrating in residential areas are not the same as safety defects. They are still a public problem. An operator may see a quiet cul-de-sac as an efficient fleet staging area. Residents may see it as their neighborhood being turned into free infrastructure.

That calls for different solutions, such as empty-mile fees, staging rules, curb-management requirements, limits on where vehicles can wait between rides, or perhaps a modern version of taxi stands. We used to understand that taxis needed stands to queue. In the app-based transportation era, too much of that logic disappeared into software. Robotaxis may force cities to remember that digital dispatch still has physical consequences.

None of this works without data.

Cities don’t need access to every proprietary detail of an automated driving system. They do need enough to govern responsibly. At a minimum, cities need visibility into safety-relevant events involving these vehicles. They also need data on remote assistance events, operating-domain changes, weather restrictions, deadheading, vehicle miles traveled, and the timing and substance of corrective actions, so independent analysts can assess whether they worked.

Without that information, local officials are left to react to headlines, company assurances, and public frustration. That is not fair to cities or to responsible operators. If an operator is performing well, transparent evidence should help it earn the right to grow. If it’s not, the public shouldn’t have to wait for a serious event before officials have the information to act.

California’s AB 1777, which allows officers to issue notices of noncompliance to driverless vehicles, is a useful step. It is still a limited tool. A traffic citation may matter to an individual driver. A notice-based system may not have the same impact on a large technology company unless it is tied to broader corrective authority.

Temporary service restrictions, whether for an hour, a day, a week, or a specific operating condition, carry more weight when the problem is recurring fleet behavior rather than a single incident.

Too Big to Pause

Scale creates leverage. As deployments grow, cities may come to depend on them for daily mobility. That makes it politically harder to slow or pause service once it is in motion. The burden lands fastest on the residents who rely on it most. No city should let automated vehicles become so central to daily mobility that their deployment becomes too big to pause before the governance system is mature enough to manage it. Checks and balances shouldn’t be there to stop the technology. They need to be there to ensure the system can endure as deployment grows.

That requires conditional expansion, shared evidence, real local authority when safety risks aren’t being addressed, and collaboration between operators, regulators, emergency responders, insurers, disability advocates, labor representatives, researchers, and residents who all see different parts of the system. A serious public process needs to bring all of those perspectives together before the politics harden.

The worst outcome would be a model that is too loose until something goes wrong, then too rigid after public trust collapses.

The vehicle may be automated. The system around it is not.

The next phase of this debate shouldn’t remain framed as innovation versus regulation. The harder choice is between unmanaged deployment and governed learning.

Somewhere, a mayor or governor is going to get this right. They won’t treat automated vehicles as a choice between obstruction and surrender. They’ll recognize the harder truth: this isn’t just a technology issue. It’s a transportation-system issue.

The leader who understands that will build a model that lets the technology advance, demands evidence when safety questions arise, and gives the public confidence that someone other than the operator is watching the system.

That leader won’t just shape transportation policy. They will help define what responsible AI governance looks like in public life.

—Guest columnist Bryan Reimer is a research scientist at MIT and co-author of How to Make AI Useful. His work focuses on how artificial intelligence and automation shape real-world systems, and he advises industry and policymakers on their deployment. This piece is co-posted by Reimer on LinkedIn.

Related stories:

Junko Yoshida, with her customary incision, hewed into this tangle of questions in a discussion with two key executives at NXP Semiconductors: Ravi Annavajjhala, vice president and general manager of the NPU group, and Nico Lehment, systems architect for the robotics segment.

The discussion began with definitions.

Lehment went with concision: “physical AI is about expressing intelligence in the physical world.”

He went on to contrast physical AI, which must learn such skills as picking up a glass from a desk, from algorithmic approaches, which require programmers to solve the challenge themselves and then code the algorithms they have developed.

Annavajjhala filled in some history, tracing a thread. Early convolutional neural networks first showed that learning networks could achieve far better results on image recognition than could human-invented algorithms. Later came generative AI, such as LLMs, that could produce often-credible text or images from a user prompt.

Physical AI takes the next step, training an AI model to interact directly with the physical world, where there can be no tolerance for error.

So what?

The NXP executives positioned physical AI as the key to a wide range of systems that could change the world. On the human-centric side, Annavajjhala envisioned humanoid robots that could be trusted companions and even caregivers for the growing elderly population of the world.

On the industrial side, Lehment saw a vast array of industrial uses for devices—not necessarily resembling humans in the least—based on physical AI rather than on preconceived algorithms. In the near future, he projected, we would see generative AI itself designing such devices. The results, he suggested might not look like anything we can imagine today.

What about the AI models?

In vehicle automation, which is arguably the most advanced instance of physical AI today, there has been a recent fascination with end-to-end (E2E) models. In this approach, a giant, exhaustively trained model takes input directly from the system’s sensors, and connects directly to the system’s actuators. The entire control system becomes a huge, inscrutable black box.

These E2E models have shown some ability to control an autonomous vehicle. But our NXP experts pointed out two categories of problems that have emerged. First, it has proved extremely difficult to implement the kinds of functional safety and security requirements common in industrial environments in an E2E network, simply by training the network on examples that illustrate safety rules. You can get an E2E network to usually follow the rules, but you can’t prove that it must follow the rules.

Second, retraining the E2E network is a huge burden, and it must be done every time you discover something new that needs to be included in the system’s behavior.

Further, there is no guarantee that the retraining to accommodate new training data will not cause new, undesirable behavior in another area. Stability of generative networks is still an area of active investigation.

Modular systems—hybrid architectures that may include several different neural networks and perhaps algorithmic modules as well—address both of these issues. Functional safety can be implemented in a supervisory block that is based on defined rules, and can employ the two-channel redundant architectures used in functional safety systems today. Similarly, a separate, verifiable block can implement vital system security functions. These two areas, the executives pointed out, draw on a rich vein of NXP’s corporate heritage.

Modularity also has architectural benefits, Lehment observed. For example, in a robot that includes sensors and manipulators, all the challenges and device specifics of sensor processing can be confined in a sensor-processing module. Thus there is no need to retrain the main neural network when a sensor gets swapped out or ages. Similarly, physical control of manipulators can be confined in a module, so the main network does not need to deal with detailed changes in the physics or control logic of arms and fingers. In this way latency between sensor input and manipulator response can be controlled, as well.

This sort of modularity opens another range of possibilities, Annavajjhala suggested. While today most systems rely on retraining their vast generative-AI models in equally vast datacenters, as infrequently as possible, small AI modules might be more easily trained: even trained in place at the edge. Some models, for example, might employ reinforcement learning to continually refine their behavior with direct feedback from the physical world. This, in turn, suggests that similar robots would need to communicate with each other, so as to share their remote learning and to stay congruent and predictable during updates.

Modularity can extend into the hardware as well. A system assembled from a variety of different modules, perhaps employing very different kinds of AI models or rules-based subsystems, will have a wide variety of computing needs. This may lead to the computing resources of the system becoming an heterogeneous network of computing elements, rather than an array of identical GPUs or AI accelerators.

Emerging platforms

What is the role of a platform vendor in this environment? The NXP experts suggested that it must go beyond just providing computing hardware and foundational software such as operating systems. Lehment warned that physical AI included some extraordinarily hard problems. The kinds of world models being discussed today can be hugely complex, and can only be assembled by training networks by experience of physical laws, not just by watching videos.

And functional safety is a relatively mature field, but its standards are extremely demanding, and were formulated, Lehment emphasized, to be usable by 50- or 100-person teams of safety engineers with long development schedules—not as just one part of the project for a dozen designers in an AI startup. Meeting a standard like ISO 15066 would be a huge ask for a start-up team.

Similarly, there are many subtle details in sensor processing and in positioning systems that will not easily be incorporated into a machine-learning network, but that are well known to experienced industrial robotics designers.

The vendor’s role

The role of a platform provider like NXP, then, has three parts, Annavajjhala said. First, certainly, is to provide hardware with the computing efficiency, in both cost and power consumption, to meet the design’s needs. But equally important is for the vendor to stay on the leading edge of model development for physical AI, and offer to clients SDKs incorporating the latest models in their domain. Finally, a vendor such as NXP can bring their decades of industrial control experience—in sensor processing, positioning and manipulation, in functional safety, and in system security—to the table as modules within the SDK.

This offering is not intended as a turnkey system in a box, but rather as a set of integrated, but separately alterable, reference designs. The object is for even the small AI startup to start from a working platform and integrate in their own new ideas. This shifts the heavy lifting for the very big and very real challenges of physical AI onto the platform vendor and frees the innovator to innovate from a known good base. It is a path for directing a fragmented, venture-funded industry forward instead of into chaos.

Links below point to the full episode on YouTube.

Chapters:

(00:00) Intro
(04:53) How NXP defines Physical AI
(07:07) Evolution of AI models
(09:27) Two camps defining ‘World Model’
(12:47) How do you embed physics into the World Model
(18:08) Roles End-to-End learning plays in Physical AI
(24:52) Do devices at the edge learn in Physical AI?
(29:27) Ensuring safety at the edge
(45:52) Seeking modularity in Physical AI

More on Physical AI: Playlist on Junko’s Talk to Us YouTube Channel

Editor’s note: Ron Wilson, a former colleague of Junko Yoshida when they both worked at EE Times, is today an independent technology analyst, writer and contributor to both Junko’s Tech Probe on Substack and Junko’s Talk to Us YouTube channel.

An open letter, co-authored by pro-AV physicians Jonathan Slotkin and Eric Topol, was a tonic for the autonomous vehicle industry. Until this endorsement, the medical community has not been writing editorials about road deaths, although these tragedies are a medical crisis that has plagued the United States for decades.

But timing matters. So, why now?

Let’s focus first on legislative issues subtly pressed in the doctors’ letter. The quandary of federal vs. state regulatory jurisdiction comes down, in short, what roles should local governments be permitted to play in regulating AVs?

Slotkin and Topol make their position clear, singling out, as villains, “New York, Massachusetts, Washington, Washington, D.C. and Minnesota” as “attempting to, or have effectively blocked autonomous vehicles despite this growing body of evidence.”

They wrote, “Whether an American gets seriously injured or worse should not depend on which side of a state line they drive on.”

Could the impetus of the open letter be more political than they led us to believe? By claiming that these states are blocking the urgent need for AV deployment, they seem to be proposing a unified national AV framework.

Second, the letter is about AV safety data. Citing Waymo numbers, Slotkin and Topol conclude that “safety data [of autonomous vehicles] has reached a threshold we cannot responsibly ignore.”

The doctors’ diagnosis is that delaying AV deployment will kill more people. There are, however, those who think Slotkin and Topol are jumping the gun. The debate now hinges on data, both available and absent, that can prove AV safety.

Federal vs. State regulations

First, let’s consider regulations.

While technology advancements matter, equally important is the regulatory environment that scrutinizes them.

MIT scientist Bryan Reimer sees the first imperative is to clarify “the roles of the federal and local governments.”

In his view, “The federal government should lead on vehicle safety, cybersecurity, national security, technical reliability, and core reporting requirements.” Elsewhere, “Local governments need authority and tools to manage curb access, emergency response, service rules, taxation, congestion, and street-level operations.”

Industry organizations and academics with whom I talked expressed concerns over the open letter, because it blames states and local communities for slowing AV deployment.

William Wallace, director of Safety Advocacy at Consumer Reports, said that every major AV developer “should stop pressing for an overly permissive regulatory framework, particularly one that would hinder states or localities from taking the steps necessary to keep their communities safe.”

Eric Teoh, director of Statistical Services at IIHS, agreed. “Serious discourse,” he said, is undermined “when jurisdictions are criticized simply for being slow to allow AV deployment.”

He went on, “This [AV] is a new thing with a lot of unknowns, and we must remember that communities know their own needs. Discourse on how to design and distribute regulatory oversight for AV deployments is healthy and should result in stronger systems.”

Michael Brooks, executive director at the Center for Auto Safety, is skeptical of the players in the AV industry, “who have accelerated their hype and hyperbole in hopes of achieving industry-favorable results in Congress and in state legislatures.”

As Brooks sees it, “Federal regulations are needed to establish minimum standards to help ensure safe AV performance. Current loose guidelines supported by AV companies have no teeth to ensure that bad actors are excluded from participation.” In parallel, he noted, “State and local concerns on safety and other impacts should be resolved rather than brushed to the side by industry-supported preemption schemes.”

Challenges to AV deployment go far beyond technology. An attorney by training, Brooks explained, “Cities and states continue to struggle to address negative impacts [of AVs], as the law around computer driver responsibility and liability is years behind at the state and federal level.”

Proving AV safety in absolute numbers

A lively discourse unfolded after I posted part one of my story (“AV: Where Did the Docs Go Wrong?”). The controversy suggests that the industry and academia are still struggling for a way to prove AV safety.

Indeed, discussions on AV safety have been confusing and divisive, largely because there is no uniformity in how the data is collected, compared and interpreted by various stakeholders.

The one point on which almost everyone seems to agree is the dire need for fair and accurate data.

Slotkin and Topol wrote:

“We call on Congress and NHTSA to establish enhanced federal data-reporting requirements for all autonomous vehicle operators: crash rates per mile, geographic deployment data, validated benchmarks for comparison, and independent safety audits verified against police reports, insurance claims, and privacy-protected medical records.”

No argument there.

However, before dissecting the AV safety data, we need to step back and consider a bigger picture.

Phil Koopman, professor emeritus at Carnegie Mellon University, for example, sees danger in proposing to prove safety with a single number – by dividing the number of crashes by miles driven.

“We are still figuring out what metrics to use,” said Koopman, “while gathering data to determine whether AVs can be acceptably safe.”

So far, small fleets show promise, Koopman said. “But as fleets scale up, safety issues are becoming more common. There is a lot of work remaining to mature AV technology for operation at scale.”

Simply put, “Net statistical harm is only one aspect of acceptable safety,” stressed Koopman. “But the AV industry has yet to acknowledge that things such as driving past school buses are also legitimate safety concerns,” he added.

IIHS’s Teoh noted, “The focus must be on understanding the safety of AV deployments and expansions, and this requires collecting the appropriate data and maintaining that data in a form that can be used effectively.”

Waymo deserves credit for voluntarily reporting vehicle miles traveled (VMT) during driverless operation. Advocates of regulation, including IIHS’s Teoh, would like to see all companies “required to report VMT in a standardized fashion.”

But looking closely at Waymo’s AV safety data, some academics say it’s too soon for the medical community to conclude that Waymo has proven its AV safety in its absolute numbers.

The question of how to perform an apple-to-apple comparison between human-driven conventional driving and automated driving is still very much a work in progress.
This is why Bryant Walker-Smith, associate professor, at the School of Law and (by courtesy) School of Engineering at University of South Carolina, regards Waymo’s analyses problematic.

He observed that Waymo “necessarily includes numerous debatable assumptions” — either by “adjusting conventional crash data so they are as comparable as possible to automated driving crash data” or by “construing conventional crash data in a way that tends to overestimate the safety of conventional driving.”

While commending Waymo’s efforts to publicize its work, Walker-Smith discerned devils in the data’s details. He said. “People tend to think that numbers are inherently objective, but those of us who work with numbers understand how imperfect and imprecise they really are. Certainty and validity are not synonyms.”

The analysts in academia worry about uncertainty that scaled AV deployment is already bringing to the real world. The industry may need to reconsider its initial efforts to measure AV safety solely in crash rates per mile.

Walker-Smith offered an example. “Until a few months ago, we didn’t fully know how Waymo performs (or performed) during a blackout.”

Waymo’s safety claims are limited to a subset of driving conditions. “Their numbers tend to measure routine driving rather than the long tail of edge cases that could have significant safety implications,” Walker-Smith said.

The same caveat applies to the use of remote operators. The impact of the AV companies’ dependence on remote operators was not foreseen before by general pucblic, nor has it yet been fully examined by the industry.

Missy Cummings, professor at George Mason University and director of the Mason Autonomy and Robotics Center, stressed the difficulty for AVs to scale safely.

“AVs don’t know anything, don’t understand anything, don’t have real reasoning. If cars can’t figure out a small puddle from a huge, deep puddle on a road, that is going to continue to be a problem.”

“Companies keep thinking that remote operations are going to help them, and it can, and they should,” Cummings said. But “there’s also a limit. All these problems are just showing the limits of remote operations. Unless your car signals that they might need a remote operator — and when the remote operator is not brought into the reasoning loop — the cars are still going to make a mistake.”

Absent crashes, puddles and power outages don’t appear in the AV safety data shared by Waymo. But if floods and lightning are not AV safety risks, what is?

Technology is making rapid progress. So is AV safety data. The status quo will never settle down, “especially if the miles being counted span significant changes in software and operational conditions,” said Koopman.

Koopman explained, “Waymo’s software has changed dramatically since the 2024 publication of a peer reviewed journal article, “Comparative safety performance of autonomous- and human drivers: A real-world case study of the Waymo Driver.”

“Waymo today is adopting a different machine learning approach, invalidating previous data. Moreover, data from those older cities does not necessarily expand to new cities with new hazards,” he noted.

Given the complexity of real-world incidents, crash data isn’t enough to capture the state of AV safety. The points on Slotkin and Topol’s wish list — crash rates per mile, geographic deployment data, validated benchmarks for comparison, and independent safety audits verified against police reports — are all important. But, according to MIT’s Bryan Reimer, “I would go further. We need reporting on exposure, operational design domains, remote assistance, emergency interactions, failures, blocked roadways, interventions, software changes, and performance over time.”

What next?

Neither the industry groups nor the professors are advocating accelerated AV deployment.

Greg Brannon, AAA’s director of automotive engineering research, summed it up. He hopes to see “continued safe and slow rollout of reliable AV technologies.”

IIHS’ Teoh said, “We need data systems in place that enable us to quickly monitor the safety of driverless AV deployments and expansions to ensure … the safety goals we all want.”

Cummings had two items in her bucket list. First and foremost, she wants remote assistance to be treated as a professional operation in which “the federal government would step in and start setting mandatory reporting, like we do for air traffic controllers.”

Second, she cited the problem with end-to-end (E2E) learning. “There is virtually zero research in academia that addresses E2E. This is largely because it’s expensive. You need to have a lot of GPU clusters. So, we have no idea of the limits of E2E learning, or how we can better improve that.”

Corporations are free to advance their commercial claims. But public needs to be informed that some of the AVs are using technology whose inner workings remain a mystery to the experts and regulators who are tasked with protecting the public from the technology’s misuse.

In short, as Walker-Smith noted, “Companies that claim that they have ‘solved the technology’ (or the like) are wrong.”

He said, “ The biggest obstacle to mass deployment remains techno-economic: How to affordably and reliably deploy demonstrably safe automated driving over a full range of driving conditions.”

Is it too much to ask the doctors to unsheathe their otoscopes and take another look?

Bottom line

In the Part 2, we have learned:

• Industry organizations and academics are concerned with the legislative issues pressed in the medical community’s open letter. They feel that state and local concerns are being brushed to the side.

• Safety experts see danger in proposing to prove safety with an overly simplistic numeric approach — by dividing the number of crashes by miles driven. They see that net statistical harm is only one aspect of acceptable safety.

• The analysts in academia worry about uncertainty that scaled AV deployment is alreadying bringing to the real world.

• The industry and academia agree on the need for a uniform way of collecting appropriate AV safety data, and maintaining that data in a form that can be used effectively.

• Neither the industry groups nor the professors are advocating accelerated AV deployment.

• The impact of remote operations and E2E machine learning on AVs should not be downplayed.

Related story:

The pressure in the tech and automotive communities to pick a side—pro-AV or anti-AV—intensified last week with the publication of “An Open Letter from the Medical Community on America’s Road Deaths.”

Citing the more than 39,000 fatalities on American roads in 2024, medical doctors Jonathan Slotkin and Eric Topol, co-authors of the open letter, emphatically declared that “safety data [of autonomous vehicles] has reached a threshold we cannot responsibly ignore.”

Well, has it?

The doctors stress the urgency for society to prioritize autonomous vehicles, to prevent much of the suffering caused by deaths and injuries on the highway.

20 doctors including the two co-authors signed the letter as founding signatories. Working under the presumption that autonomous vehicles are inevitable, they argue that we should all get onboard to save people’s lives.

Looking more closely, however, it’s unclear whether currently available data or public opinion support this presumption.

The doctors hinge their argument on the following claims. Some key aspects of the letter are:
- No other cause of death at the scale (of traffic fatalities) would be met with so little urgency.
- Autonomous vehicles safety data has reached a threshold that the medical community cannot responsibly ignore.
- Waymo is the only AV company publishing the data clinicians need to evaluate health impacts.
Their call to action is that states and municipalities should not resist AV companies’ efforts to deploy.

In other words, the two doctors seem to be saying that AVs are a miracle cure for road fatalities, and imply that any delay in scaling up the technology would be as irresponsible as withholding a miracle cure for fatal diseases. But does the data they refer to really support such a bold conclusion?

While bowing to the doctors’ concerns about road deaths, it’s prudent to cast a wider net, examining the perception of AVs among all the stakeholders who might be affected.

For this article, I turned to executives at the American Automobile Association (AAA), the Insurance Institute for Highway Safety (IIHS), Consumer Reports, scientists, engineering professors, lawyers and industry observers who regularly speak about transportation and urban issues.

The result is a two-part analysis of where industry experts believe autonomous vehicles stand today (part 1) and what must happen next to push AV safety on the technical and regulatory fronts (part 2).

This article, Part 1, won’t include a rehash of the AV aspirations touted by robotaxi operators. Such claims have been amply quoted in the press and regurgitated by Doctors Slotkin and Topol.

Waymo co-CEO Tekedra Mawakana has told “60 Minutes” that Waymo’s driverless cars are “five times safer than human drivers.” This bold assertion, effective on TV, doesn’t really help to assess where autonomous vehicles stand today.

So, let’s get to work.

Is AV a miracle cure for road fatalities?

MIT scientist Bryan Reimer has posited that “AVs should be treated as neither a miracle cure nor a menace.”

His objection to the Slotkin/Topol letter lies in its framing of the discourse on autonomous vehicles. “The public conversation cannot be reduced to a binary choice between embracing AVs or blocking innovation. The real question is how to make vehicle automation a trustworthy part of a broader safety system.”

The co-authors’ apparent lack of background knowledge on transportation issues is the obvious chink in their credibility.

Missy Cummings, professor at George Mason University, and director of the Mason Autonomy and Robotics Center, characterized the doctors as studious amateurs. She told me, “Just because I am well read about medicine doesn’t make me an expert in medicine.”

It’s likely that the open letter’s signers, including the co-authors, were unaware of the role that remote operators play in operating robotaxis, and the uncertainty it brings to public roads, Cummings said.

The brittleness of autonomous vehicles is real, she stressed. “We see every day in the news stories” on AV mishaps that illustrate her point.

“Even if AVs can reduce crashes (which is possible) they certainly aren’t the most reliable, cost-effective or quickest way to do so,” pointed out David Zipper, a frequent commentator on transportation, cities, and technology. “Topol and Slotkin may be experts in venture capital and medicine, but they lack the background in transportation that would enable them to compare AVs against other forms of road safety interventions.”

Although “road deaths are a public-health crisis,” Reimer pointed out, “no single technology gets to declare itself the cure without transparent evidence, independent scrutiny, and public accountability.”

My first question to my sources was this:

Is AV the best way to reduce road deaths? If not, what other solutions would you suggest?

Many experts agreed that autonomous vehicles have “potential in the long term to improve safety.” None believes this potential is imminent.

Bryant Walker Smith, associate professor, at the School of Law and (by courtesy) School of Engineering at University of South Carolina, said, “Automated driving could help, if we’re careful about it. But people are dying today not because we’re careful about automated driving but, rather, because we’re careless about road safety generally.”

William Wallace, Consumer Reports’ director of safety advocacy, sees “the enormous potential of AVs,” but made clear that AVs can save lives only in the long run, “if they are developed and rolled out responsibly – over 10, 20 or 50 years.”

The US is an outlier

For evidence of AVs saving lives, it’s helpful to look beyond the United States. Wallace called the US “a global outlier,” citing other nations that have already reduced road injuries and fatalities substantially without depending on AVs.

Phil Koopman, professor emeritus at Carnegie Mellon University, noted, “Europe has already shown us that we can do dramatically better on road deaths without autonomous vehicle technology.”

He suggested, “For example, if we handled drunk driving more seriously as a public health concern, we might cut those fatalities in half to catch up to where the UK and many other countries have been for decades.

Impaired driving isn’t the only cause.

Citing a vast body of research, Cummings was unequivocal: “Speed kills people on the road. If we could get speeding under control, many, many, many, many, fewer people would die. This doesn’t have anything to do with self-driving.”

How long would it take?

In their letter, Slotkin and Topol wrote that “more than 100 Americans die in car crashes every day.” They said, “No other cause of death at this scale would be met with this little urgency. As clinicians, we no longer accept it.”

By the doctors’ metric, the best way to remedy this unacceptable crisis is to start with the proliferation of AVs. Indeed, the standard pro-AV argument notes that AVs eliminate such major causes of crashes as speeding, impaired driving and distraction.

However, Greg Brannon, AAA’s director of automotive engineering research, who agrees with the promise of AVs, warned, “With over 280 million vehicles on the road, it will take decades before we see significant benefits from the technologies.”

Eric Teoh, director of Statistical Services at IIHS, echoed a similar sentiment. He said, “Driverless AVs certainly can play a part, but doing so is not as simple as accelerating their deployment.”

IIHS has been monitoring rideshare services for two decades. Despite their popular acceptance, rideshares have yet to divert people from driving their own cars. “There is no reason to expect that a major shift to robotaxi will happen,” Teoh deduced.

In his opinion, “AVs’ lower crash rates will only translate into fatality reductions if they replace riskier travel, rather than just adding more [AV] miles on our roads.”

On the issue of AV market penetration, the geographical divide is significant. Michael Brooks, executive director at the Center for Auto Safety, observed, “Due to lower demand and rural connectivity and infrastructure issues, AVs will not be deployed in rural areas for decades.” This matters because “rural crashes represent a large proportion (~40%) of road fatalities and injuries,” he pointed out. “AVs are not going to address these areas nearly as quickly as implementation of far less expensive safety technologies in human-driven vehicles.”

Brooks also noted the time it takes for individual safety technologies to fully penetrate the vehicle fleet. Typically, this is about 30 years. Replacing the entire fleet with autonomous vehicles, however, is not likely to be typical. It will take far longer. “NHTSA and other federal agencies should be pushing industry to adopt the near-term lifesaving solutions rather than relying on the AV industry for safety gains over the next few decades,” he said.

The promise of AVs has caused distraction

Many industry experts are concerned that by prioritizing AVs, we may be taking our eye off many safety solutions already proven to save lives.

Reimer noted, “The danger is treating automated vehicles as a replacement for the harder work of building a safer transportation system. For years, the promise of fully driverless vehicles has too often distracted from safety improvements that could help drivers today.”

Reimer said a more comprehensive Safe System Approach would include speed management, road design, pedestrian and cyclist protection, impaired-driving prevention, safer vehicle design, enforcement, transit access, driver assistance, and post-crash medical response.

Consumer Reports’ Wallace said, “These solutions are widely available and yet they aren’t currently getting the kind of investment they need to more effectively reduce road deaths in the short term.”

Zipper concurred. “While AVs’ safety benefits remain speculative, there are also many other ways to reduce crashes that are proven to work and are available right now – at scale, if we choose to adopt them.” Beyond the conventional Safe System Approach, Zipper’s examples include “installing protected bike lanes, deploying Intelligent Speed Assist, slowing high-speed urban arterials, building complete sidewalk networks, banning truck lifts, and providing high-quality transit and train service.” He added, “Travelers are much safer in a train or a bus than in a car.”

To be fair, Dr. Slotkin and Dr. Topol acknowledge in their open letter, “Autonomous vehicles are not a replacement” for such efforts as “safer infrastructure, better speed enforcement, impaired driving prevention, and public transit.” Nonetheless, they maintain that “AV safety data has reached a threshold we cannot responsibly ignore.”

AVs’ journey to safety

My second question to my sources was:

Where are we on the journey to safety with AVs?

As robotaxis proliferate, one issue looming large on experts’ minds is whether safety will scale in parallel with accelerated operations.

Teoh, IIHS’ director of statistical services, noted, “Driverless AVs, particularly Waymo, have been shown by several studies to have lower crash rates than human drivers … While the technology is impressive and continues to grow, it’s unclear if these benefits will remain when expanding to new cities, road types, and conditions.”

Koopman concurred. “So far things are promising for small fleets, but as fleets scale up safety issues are becoming more common.”

Contrary to “the AV industry’s premature declaration of victory on AVs,” Koopman stressed, “there is a lot of work remaining to mature AV technology for operation at scale.”

AV failure modes

Cummings disputes the casually cited notion that self-driving cars are already established as safer than human drivers. “There is no statistical evidence,” she said.

In her view, “We still have a long way to go before self-driving cars will be able to be seen at least as competent as good drivers.”

Asked about AV safety, Cummings said, “We’re not even close... We’re still learning about self-driving car failure modes every day.”

Citing a flood in Texas, which recently caused Waymo to recall its entire fleet, Cummings said, “Man, that is a canary in the coal mine. How much longer is this going to have to go on?”

Flooding is just one road hazard demonstrating that “these cars do not cope with uncertainty,” she noted. “Humans know that they’re driving into a flood. Self-driving cars don’t know anything, they don’t understand anything.”

The dilemma, Cummings noted, is that lidar can see water but cannot read it. When an AV detects water on the road, it doesn’t know if this is a puddle or ten feet deep, she explained.

Absent sensors that can detect water depth, this remains an unsolved problem, she noted. “The only way that self-driving cars can deal with that right now is to make sure that they have fleet sensing … If one car finds out about a flooded road, it sends signals to everyone else.”

Reimer pointed out, “Deployment capability [of AVs] is not the same as system-level safety readiness.” Currently, he said, “Deployment is scaling faster than public discourse and institutional capacity.”

In Part 2, we will discuss the safety data that’s available today, and what next steps — institutional and regulatory — must be taken to make AV operations safer. We also explore the possible political motives behind Doctors Slotkin and Topol’s open letter.

Bottom line

In the Part 1, we have learned:

• AVs can save lives only in the long run.

• With more than 280 million vehicles on the road, it will take decades to see significant benefits from the technologies.

• The U.S. is a global outlier. Europe has already shown how to dramatically reduce road deaths without autonomous vehicle technology.

• A comprehensive Safe System Approach already offers widely available solutions that can effectively reduce road deaths in the short term.

• Yet, the promise of fully driverless vehicles has too often distracted the industry from safety improvements that could help drivers today.

• As robotaxis proliferate, a big, unanswered question remains. Can safety scale in parallel with accelerated operations?

• The brittleness of autonomous vehicles is real. The AV industry is still learning every day about how self-driving cars can go wrong.

When my 12-year-old looked up the word autonomous, the definition was straightforward: self-governing, independent. That is exactly why the term breaks down when applied to robotaxis.

Calling them autonomous flatters the technology and hides who is responsible when something goes wrong. It shapes how the public and policymakers understand these systems. Autonomous suggests a machine that governs itself.

But the systems now being deployed on public roads, especially robotaxis, rely on human support, operational oversight, maintenance, and institutional coordination that the label obscures. Calling these vehicles autonomous does not just flatter the technology. It narrows accountability.

That matters because language shapes governance. If leaders believe they are regulating a self-sufficient machine, they may focus too narrowly on the vehicle itself: its sensors, software, and safety record. But if these systems are service networks with hidden human inputs and fragile dependencies, the policy questions become more urgent. The real question is whether policymakers are willing to regulate the whole system rather than the branding around it.

Military drones are often discussed as though they operate alone, even when they rely on remote operators, communications links, intelligence inputs and institutional rules of engagement. The machine may be distant, but the system remains deeply human.

The same is true on public roads. Robotaxis may remove the driver from the front seat, but they do not remove people from the system.

The Human Layer Behind Robotaxis

For years, the story around robotaxis was simple: cars that drive themselves, less human error, safer streets, cheaper rides.

The reality is more complicated. These systems do not run themselves. Someone answers rider questions. Someone cleans the sensors. Someone sits at a desk ready to step in when the car gets confused. There is maintenance, software updates, and mapping refinements. None of that is glamorous, but it is what keeps the service running.

This does not delegitimize the technology. If anything, it clarifies what makes it work. Redundancy and human backup may be exactly what makes these systems as safe as they are. But the public deserves a more honest accounting of what autonomous really means in practice. At that point, autonomous starts to look less like a technical description and more like a marketing choice that has stuck for too long.

The Governance Question Cities Cannot Dodge

Many local leaders are made to feel they are either pro-robotaxi or anti-innovation. That is a false choice.

The real question is whether officials understand what is being deployed on their streets and are willing to govern it, rather than simply get out of the way. Cities do not need to design the technology. But they do need to coordinate with firms to shape how the service operates in public space.

That means asking questions well beyond vehicle performance. Who provides remote assistance, and from where? What standards govern intervention? What happens when communication fails, or when flooded roads and other edge cases create uncertainty? Where do vehicles go when they are not on active rides? How should residents report unexpected or unsafe behavior? How are vehicles maintained, inspected, cleaned, and recalibrated? How do emergency responders interact with the vehicle and its remote support? How quickly can local officials obtain meaningful records after an incident?

California’s new regulations may begin to point the way, but cities and states cannot answer these questions on their own. They will need shared frameworks, coordination with technology providers and other stakeholders, and a broader view of what these systems involve. These issues are central to public safety and public trust and are too complex for individual localities to manage alone.

I remain skeptical that robotaxis will deliver the low-cost rides often promised. The layers of redundancy are expensive. That does not mean the business cannot succeed. It means profitability may be harder than many investors and innovators expect.

When profitability proves difficult, pressure to cut costs follows. Unless governance keeps pace, cost pressure will create incentives to weaken some of the support layers that currently make these systems publicly acceptable. Elected officials, not just corporate leaders, should be debating those tradeoffs.

The same logic extends beyond safety. Will these services lower transportation costs, or simply shift costs into less visible forms of labor and infrastructure? Will they complement public transportation, or erode transit ridership the way ridesharing did in many cities? Will they improve access to underserved communities? Will deadheading increase congestion? Will local communities bear the burdens on public space while companies retain most of the operational visibility?

If policymakers do not ask those questions up front, they will ask them after a crisis, when politics turn reactive, and trust is already eroding.

Federal Rules, Local Terms

Senator Markey’s scrutiny of remote assistance and offshore support has helped surface what should have been clear all along: this is not just about vehicle safety. It is also about labor, national security, transparency, and control. California’s debates over infractions, remote operators, and emergency response point in the same direction. What is really at stake is how visible and governable the system must be before the public is asked to trust it.

Federal and local governments have different roles, but both must get this right. The federal government should take the lead on vehicle safety, cybersecurity, national security, technical reliability and whether critical remote assistance can be performed offshore. Local governments, by contrast, must manage the street-level realities of deployment: service requirements, curb access, ticketing, taxation, emergency interaction, and the conditions under which these systems enter civic life.

Cities cannot punt this to Washington. They will be left managing it on the ground.

Regulating the System, Not the Story

Constructive governance does not mean blocking the technology. It means setting terms that reflect reality.

Leaders should stop treating robotaxis as if they are merely cars with better software. They are service systems that require oversight not only of the product but of the people, processes, and institutions that make it work.

That means more robust public data, not just on crashes but on exposure, interventions, operational limits, failures, and performance over time. It means clearer standards for remote support and emergency interaction. It means understanding where labor sits within the system and what happens when economic pressure makes that labor less visible. It means asking not just whether the vehicle can drive, but whether the larger system can scale without weakening safety, accountability or public confidence.

Better language leads to better questions. Call these systems what they are, highly automated and still human-dependent, and the governance questions become harder to ignore. Get that right, and long-term public trust is more likely to follow.

Robotaxis are likely to become part of the mobility landscape. The jurisdictions best positioned to benefit will not be the ones that echo the industry’s preferred vocabulary. They will be the ones willing to govern the system as it actually exists.

—Guest columnist Bryan Reimer is a research scientist at MIT and co-author of How to Make AI Useful. His work focuses on how artificial intelligence and automation shape real-world systems, and he advises industry and policymakers on their deployment. This piece is co-posted by Reimer on LinkedIn.

Other stories by Bryan Reimer:

Earlier this week I attended the Embedded Vision Summit in Santa Clara, Calif. While I enjoyed returning to my old stomping grounds in the Bay Area, I remembered how the tech community’s “bubble in a bubble” milieu felt.

Two things struck me.

First, Silicon Valley’s engineers still feel no fear of failing fast. Their optimism for their self-styled tech revolution—just around the corner— remains passionate and pervasive.

Second, nonetheless, they still struggle to anticipate what comes next—especially the unintended consequences—after the latest technology rollout.

Whether the trend is the proliferation of social media, ChatGPT or autonomous vehicles, the wizards of the Valley seem chronically oblivious to the impact of their remarkable technological feats on people and society at large. They have, after all, reaped their billions and they’re moving fast to the next big thing.

On full display at the Embedded Vision Summit was a mix of new hardware and software including vision sensors, edge processors and new AI models. All are designed to enable new “category creating” applications that range from edge, physical and agentic AI to cloud, industrial automation and autonomous vehicles.

Native on-device agent

With physical AI still a few years out (except for robotaxis), agentic AI seized the summit’s spotlight. Speakers and attendees shared their experiences with what appears to be a rapidly advancing “agentic flow,” or agents’ “harnessing” abilities that can potentially affect their business.

One particularly notable keynote was entitled “Scaling Down is New Scaling Up” by Meta’s AI researcher.

This speech piqued my interest because it shifts the conventional narrative of “agentic AI” — enterprise-level productivity tools—to an individualized, personal agent at the edge.

Keynoter Vikas Chandra, senior director at Meta Reality Labs, presented his team’s efforts to power “private contextual assistants” with sub-billion parameter reasoning models.

Translation: In plain English, he was talking about developing an AI “agent” unique to the needs of an individual, perceiving the world through that person’s senses.

The agent learns personal preferences not just through text interactions (via Chat GPT prompts, for example) and but more importantly by continuously collecting data — through smartphones, watches, glasses and other wearables. Intimately it learns how the person sees, hears, senses and reacts to the physical world.

The context that’s relevant to you as an individual is “built — persistently — over days, weeks, months, maybe even years, to exactly become an assistant that remains with you all the time,” explained Chandra.

“My thesis,” he said, “is that we are getting into a world where we’ll move away from chatbots. Maybe not completely, but to enable a personal agent that’s always present, always private, always ready.”

Bigger-the-better scaling models can be reversed, Chandra believes. The question will become “how much capability can we add to enable these [personal, edge] devices to be smart and intelligent.”

Many semiconductor companies are already looking to develop smaller, more effective AI processors for the edge. Meta’s Chandra is proposing to scale down AI “algorithmically.”

Chandra defined such a native on-device agent with abilities to:

• Reason over complex tasks

• Perceive the world through multiple modalities

• Maintain persistent context across interactions

• Operate within a strict memory and power budget

His presentation included specific technical progress his team has made with “extreme quantization” to fit an agent brain onto a phone, the “architecture” of an agent’s core, and “runtime” to make the agent’s seem instantaneous.

Keep it private?

Despite Chandra’s pledge that Meta would keep personal agents “private,” I suspect that he would face a challenge in quelling doubts. This is the same Facebook whose corporate history has demonstrated scant reverence for the privacy of its constituents.

Nonetheless, Chandra stressed that, “The most useful agent knows your life. The most trustworthy one keeps it there.”

Yes, it is a great slogan, but wouldn’t building a personal agent end up being not so different from installing a personal surveillance system?

If the individualized AI agent functions as well as advertised — functioning as your clone, knowing your every idiosyncrasy, anticipating your every move — couldn’t that seem just a little creepy? Wouldn’t you worry just a little about all this intimacy leaking out onto the world at large?

After all, until we know how Meta is going to make money on a personal surveillance device, we should be all skeptical of the comapny’s motive.

Waymo’s safety claim

Elsewhere at the summit, Waymo was strutting its stuff.

Chen Wu, Waymo’s head of “perception,” claimed that Waymo’s fleet has achieved “significant safety metrics, reducing overall crashes by 92%, serious injuries by 83%, and airbag deployments by 82%.”

Operating in 11 US cities, plus Tokyo and London, Waymo delivers more than 500,000 paid rides a week, according to Waymo.

The highlight of Chen’s preesntation was sharing real-world examples of quick Waymo reactions to potential collisions, such as stopping for a pedestrian, red light runner, and scooter falling.

Chen showed a video clip of two people, in the dark, climbing over a divider and popping up suddenly on the road, demonstrating the robotaxi’s skill at avoiding pedestrians, unexpectedly appeared in a hard-to-detect environment.

What Waymo didn’t talk about

Missing, however, in Chen’s presentation was mention of 3,800 robotaxis that had to be recalled over the risk of entering flooded roads. The recall news broke very early in the morning of last Tuesday — the same day Chen was giving her speech entitled “Building Trustworthy Autonomous Driving Systems: Waymo’s Holistic AI Safety Approach.”

Chen also elided reports about Waymo’s robotaxis driving past school buses flashing their red stoplights. Such incidents have been recorded at least 19 times. Asked for an explanation during Q&A, Chen noted, “Yeah, so actually, very recently, after we saw those instances, we made a ton of progress on school buses.” She went on to suggest that Waymo’s robotaxis have trouble discerning the difference among school buses. “There are small differences,” she said. “The stop signs are here, not here, or where the flashing light can be different.”

Acknowledging that “sometimes we learn things new that we didn’t know,” Chen said, “The good news is that the development ecosystem is able to find examples and we can use a small number of examples to quickly improve the software… so but yes, keeping anticipating new situations is something that’s important for us.”

I can’t blame companies for ducking discussion of inconvenient truths in a marketing presentation. But if “Building Trustworthy Autonomous Driving Systems” is indeed the goal, AV companies would be well-advised to shore up public trust by sharing information not only on their accomplishments but also the mysteries they’re still working to solve—like those big yellow things with the blinking red lights.

Bottom line

The Embedded Vision Summit paraded the usual stellar lineup of speakers and presentations touting nouvelle technologies. But when tech notions gravitate toward commercialization (a la robotaxis), the audience beyond the Bay Area’s propeller-head monasteries needs a reality check that goes far beyond claims, boasts and rosy promises.

The hardest challenge for automated vehicles may not be technical. It may be institutional. We’re deploying these systems faster than we’re willing to discuss them. We still have not seriously debated where these systems should operate, what they owe the public, or who pays when they fail.

Automated vehicles are only the most visible version of a much larger problem in AI governance. Congress, federal agencies, and state and local governments are being asked to make decisions about systems they often do not yet have the capacity to evaluate.

A recent Ezra Klein conversation with Alex Bores, a New York Assemblymember and congressional candidate, helped crystallize this for me. It sidestepped the usual extremes: reflexive fear of the technology on one side, cheerleading deployment on the other. Instead, it focused on the harder questions. Who benefits? Who bears the risk? Can democratic institutions help shape these systems constructively before markets lock them in?

There’s a real case for slowing down parts of the AI ecosystem. If AI is already beginning to reshape labor markets, education, policing, energy, national security, and politics, caution is not obstruction. It is a democratic responsibility.

But moving too slowly also carries real risks: strategic, economic, and geopolitical. That is the real AI tradeoff.

A broad Western slowdown could give China and other competitors a strategic advantage. This doesn’t mean throwing away guardrails. It means we can’t govern by delay alone.

China doesn’t innovate in a vacuum. It adopts, adapts, and scales. Its automotive boom should be taken seriously. Built partly on Western innovation and then accelerated by aggressive industrial policy, manufacturing scale, lower cost structures, and fast iteration, China has shown how years can compress into months. AI may compress that timeline even further. In automated vehicles, Chinese companies such as Baidu and Pony.ai are moving quickly, while Western AV companies continue to wrestle with business models, regulatory uncertainty, labor politics, and public trust.

None of this is simple.

Move too fast, and we deploy systems before society, law, and institutions have caught up. We also risk eroding skills we should not be so quick to outsource. Writing, reasoning, judgment, problem-solving, and deliberation are not luxuries. They are part of what democracies are built on.

Move too slowly, and democracies risk handing leadership to governments and companies with weaker commitments to transparency, accountability, privacy, labor, and consent.

There’s no acceptable shortcut. The challenge isn’t choosing between innovation and guardrails. It is building guardrails while the road is still being paved.

What does this look like in practice? Mandatory incident reporting before the press finds out. Independent audits before market dominance, not after. Taking labor impacts seriously before displacement is treated as inevitable. Real-time feedback loops, not investigations years after the market has already settled. Rules that evolve with the technology, not regulations written once a market leader is entrenched.

Democracy can’t govern by delay. We need what I’d call competent acceleration: moving fast enough to compete, while building the reporting, audit, standards, and oversight capacity needed to keep trust and public value from becoming afterthoughts.

Once deployment hardens into infrastructure, the choices narrow quickly.

This is why the autonomous vehicle debate matters.

A chatbot changes how we write, how we learn, how we search, code, and teach. A robotaxi changes how streets function, how emergency responders do their jobs, how cities manage curb space, who gets served and at what price, who gets displaced, and what safety means in public. These are not theoretical questions anymore.

Automated vehicle governance isn’t separate from AI governance. It’s where the abstraction stops. Real streets. Real workers. Real institutions. Real consequences. That is where fast governance has to become more than a slogan.

— Guest columnist Bryan Reimer is a research scientist at MIT and co-author of How to Make AI Useful. His work focuses on how artificial intelligence and automation shape real-world systems, and he advises industry and policymakers on their deployment. This piece is co-posted by Reimer on LinkedIn.

Other stories by Bryan Reimer:

Tips about partnerships and investments in autonomous vehicles (AV) pop up and get touted every week, if not daily.

The resulting deals involve semiconductor suppliers, traditional OEMs, AV software developers and service providers.

All this churn makes the arrival of commercial operations of AVs seem inevitable, on public roads in cities around the globe. Clearly, everyone is scrambling to catch up with Waymo, which boasts 200 million miles of driving experience, and with Waymo rivals, Zoox and Tesla.

The AV dealmaking club includes chip suppliers like Qualcomm, Nvidia, and Mobileye and automakers such as Volkswagen, Hyundai, Mercedes-Benz, Nissan, GM, Toyota and a slew of Chinese OEMs.

Wayve, a London-based AV software stack developer, sits in many deals, promoting end-to-end machine learning for AVs. Among service providers, Uber is doing its darnedest to shoe-horn into just about every recent AV deal.

The deals, nonetheless, might be just feelers, as the players size up their partners and build their AV supply chains. Juan Carlos Aguilera wrote recently in Substack, “The AV industry’s value chain doesn’t exist yet. Companies are assembling it in real time, and the most honest signal is how often they change roles.”

In parallel, there are telltale signs of emboldened geopolitics playing out, in search of AI sovereignty. The Japanese government, for example, has given a directive to its auto industry to achieve “a global market share of 26 percent in cars with autonomous driving technology in the 2030s,” according to a report by Japan’s public broadcaster NHK.

As geopolitics rears its head, companies tend to root for the home team. Silicon Valley vs. London was a theme expressed by James R in his Substack piece, “London Calling - for the Robotaxi revolution.” He writes: “If Wayve wins in London, that’s a British company beating Google on British streets. It’s a bit galling that the UK had the talent to lead this race and instead we’re welcoming in American companies to go first, but at least we’re in it now.”

Instead of speculating on the next big AV deals, I propose a pause. The media and regulators would be wise to crack a history book and look back 100 years.

Why bring this up? Some of what we perceive as new challenges facing the AV industry echo the experience of the rail industry in the era of robber barons.

Cornfield Meets

Challenges dogging many AV startups today are the “speed” of technology development and the “scale” of business operations. Fighting for survival, every company rushes to hit milestones set by investors. For the investment community, speed and the scale, not safety or cost per vehicle, are the priorities.

Looking back, the United States witnessed an explosion of railroad construction between 1860 and 1900. The rail industry’s ambitions were shaped by rapid innovation in steel and steam. The railroad tycoons, in relentless pursuit of wealth and influence, adopted a “move fast and break things” ethos that foreshadowed tech tycoons like Alphabet and Amazon in pursuit of the autonomy economy.

The rail industry’s growth was extraordinary. Prior to 1871, approximately 45,000 miles of track had been laid. Between 1871 and 1900, another 170,000 miles expanded the transcontinental railroad system.

But one consequence of this massive scaling — with many more trains in operation — was an increased number of crashes called “cornfield meets,” explained Phil Koopman, safety expert who was in our recent podcast with our historian Peter Norton.

“Cornfield”?

A cornfield meet is a head-on collision of two trains in rural, open country — often alongside a cornfield. In those days, to save money, trains going in opposite directions often shared a single track. When trains got their timing wrong or the telegraph line was down, the eastbound locomotive was in danger of meeting the western express, head-on, in a cornfield. Koopman explained, by scaling up train traffic without more tracks, the railroad companies scaled up catastrophic crashes.

A consequence of all these train wrecks was the birth of Standard Time, said historian Peter Norton. It enabled rail companies to synch train schedules. The industry also created block safety systems, designed to restrict sections of track to one train at a time, added Koopman.

These safety measures were only enacted after a rash of tragedies. Initial safety measures worked well enough at first, but started failing when the industry scaled up operations.

Sound familiar?

A popular myth, debunked by history, is that Silicon Valley invented strategies like “move fast, break things,” and “ask forgiveness, not permission.” Because we live in a time of extraordinary tech innovations, the argument goes, tech companies must be allowed extraordinary technology development principles. Examples include Minimum Viable Product, Continuous feedback loops and Data-driven decisions.

In short, let’s get the technology on the road right away, and see what happens. Eventually, the data will reveal if there are cornfield meets, as well how many, how often and who got killed. We’re talking business here. It’s not about perfection. It’s all about iteration.

Getting back to robber barons and their trains, we can see clearly the recklessness of the behavior and business principles, as they aggressively avoided regulations and interpreted the laws to their benefit. Their echo is the ethos embraced by Meta, OpenAI, Amazon, Tesla, and Alphabet.

Back to the AV race in 2026

Today, if you’re an AV startup challenging Waymo, you have a couple of options.

You could, for example, tout differentiation, claiming, noted Norton, “We are not Waymo. We’re doing something quite different here”

He added, “But anybody who wants to be like Waymo at this point is certainly going to have to take risks. Otherwise they’ll be left behind.”

London-based Wayve sees this as an opportune moment, as it pitches its E2E-based AI learning system as a perfect solution for Waymo wannabes to achieve the necessary “speed” and “scale.”

Koopman explained, “as far as I can tell,” E2E enables a much smaller AV dev team to design, far more quickly, autonomous vehicles that are proficient and competent in driving.

“But it’s much harder to teach AVs edge cases,” he cautioned. Under E2E, “You can’t tell just the AV, here’s what to do. [Instead] You must give it a bunch of training data, and it must statistically infer what the action is in that particular situation,”

The school bus case

In his opinion, this is a possible explanation for problems with robotaxis rolling past school buses with children present.

To an AV a stop sign mounted on a pole is entirely different from a stop sign—and red flashing lights—on a school bus.

Koopman explained. “Stop signs and red flashing lights mean stop and then go, unless they’re on a school bus, in which case it means stop and then stop.”

Koopman acknowledged that he doesn’t know if this stop-sign/go-light quandary is the root cause of the school bus incidents, but he’s sure of one thing. “This is the kind of problem I would expect to be difficult to solve in [E2E] machine learning.”

Bottom line

If reporters, influencers and investors are eager to reward any AV company that gets 100 cars quickly on the road in their effort to catch up with Waymo, E2E can turn the trick. The perception will be that the development team has accomplished 99 percent of its goal.

The hitch, however, is the one-in-a-million problem, said Koopman. With a million—not a hundred—cars on the road, the one-in-a-million catastrophe “is going to happen every day.”

Did the rail industry think about the potential increase of cornfield meets when it was scaling up?

Most likely not.

But AVs in the United States? Tech companies can still focus on preventing their own cornfield meets or, if they can’t figure that out, scaling back ’til they do.

Related stories:

Phil & Junko on AV Safety

Podcast: The Evolution of "Safety First" with Peter Norton

Safety First with Peter Norton (Phil & Junko on AV Safety ep02…

Listen now

3 months ago · 6 likes · Phil & Junko on AV Safety

Last week posed a dilemma for mavens of autonomous vehicle (AV) development and deployment of commercial autonomous vehicles (AV). They had to choose between two AV-related conferences, the Ride AI Conference in San Francisco and, in New York, the “Future of Transportation Conference.”

Differences between the two, each with its own mission, audience and orientation, couldn’t have been starker. They offered a microcosm of today’s divided public discourse on robotaxis, especially now that AVs are roaming on public roads in many cities in the U.S.

The Ride AI Conference stole headlines by trotting out a Who’s Who roster of “the architects of the autonomy economy,” as the organizer described.

The Future of Transportation Conference at Roosevelt House, a public policy institute at Hunter College in New York City, heralded a tagline of “Autonomous Vehicles: The Benefits, the Risks, and the Need for Regulation.” While the event included representatives from the AV industry, the conference allowed dissenting opinions on AV deployment by leading thinkers of public policy on transportation.

In sum, the New York event offered an antidote to Silicon Valley tech zealots who see regulation as an anti-business principle that stifles innovation.

The following story highlights the second panel, focused on the challenges AVs still face.

A panel at the New York conference, “Autonomous Vehicles and the Hard Problems — Safety, Streets, and Tradeoffs,” consisted of Peter Norton, associate professor at the University of Virginia; Rachel Weinberger, the Peter W. Herman Chair for Transportation at the Regional Plan Association; and David Zipper, contributing writer at Bloomberg and cohost of “Look Both Ways with David and Wes.” Moderating the discussion was Sam Schwartz, Chair of the transportation research program at Roosevelt House.

Unlike most tech conferences built on the next big high-tech thing and promises of revolutionary technology, the panel urged the audience to ponder what history has taught.

Key questions discussed on the panel include:

• What “problems” are autonomous vehicles designed to solve?

• To whom do public roads belong?

• What prompted New York to hit a pause on AV deployment?

• Have policymakers and regulators shirked the responsibility to think ahead about the impact of AVs on public street.

What problems do we want AVs to solve?

Every AV company’s marketing mantra is that it’s all about safety. They’re developing this technology to save lives.

So, was safety Waymo’s preeminent motive for coming to New York? David Zipper remains skeptical. The number of people—205—killed in crashes in New York City last year, he said, “was the lowest number since 1910.” The city pulled this off without relying on robotaxis.

Zipper concedes that AVs might be an answer. But he said, “I don’t know what the question is.”

Rachel Weinberger agreed, calling AVs a “solution to a problem that’s not defined.”

In Zipper’s opinion, AV technologies might be more effective outside the city. He said, “There are so many single car crashes that take place in rural areas. There’s no real transit and there’s no real ride hail service.’ A Waymo shift to country roads, he stressed, would be “an incredible boost, not just of safety, but of mobility.”

Are any AV companies interested in wide-open spaces? Zipper wishes they were but sees no evidence.

Weinberger goes a step further. In her opinion, the crux of the matter is that “we haven’t decided what we want” from AVs and AV companies.

She noted:

“We’re allowing a new technology that’s driven by the logic of capitalism to push into our city and with an expectation that we’ll accommodate.”

Who owns public roads

It’s hardly radical to assert that public roads belong to all stakeholders, including pedestrians, cyclists and automobiles. But Peter Norton, the panel’s historian, offered the reminder that it is not always the case, especially when business interests define the rules of the road.

A case in point happened in New York in November 1922. The city’s traffic magistrate, William McAdoo, was appalled at the toll of New Yorkers, especially children, being run down, injured or killed every day by automobiles. He said, “We’re going to have to remind motorists that the streets do not belong to motorists, and we’re going to begin by banning pleasure cars south of 14th Street.”

That statement triggered “a panic attack” among automobile interests, explained Norton. Within days, a road builder writing in the Engineering News Record said, “The obvious solution lies only in a radical revision of our conception of what a city street is for.”

Provoked by McAdoo’s proposed car ban, the auto industry resolved that “we need to somehow get to a place where people think of streets as primarily for cars,” according to Norton.

Driven by a powerful industry whose wealth depended on selling cars and building roads, many U.S. cities in the early 20th century ended up devising an approach to safety simply by getting pedestrians out of the way on roads largely reserved to automobiles..

The case of New York

Waymo quietly stopped testing its AVs in New York when the city discontinued its operating permit on March 31. Gov. Kathy Hochul also tabled legislation that would have opened all of New York state to driverless vehicle services.

Sam Schwartz asked the panelists whether New York City poses issues for AV operations not present in cities like Austin and San Francisco, where robotaxis are allowed. He asked, in sum, whether the governor made the right decision?

Weinberger noted, “We are apparently different from others, but that has more to do with political pressures.” New York’s hesitance represents an opportunity for the public to “really think about what we want from AVs.” Rather than deciding how to make autonomous vehicles work in the city, she stressed, we should be considering “how can we use autonomous vehicles to do the thing that we want—when we can figure out what we want?”

Norton sees Governor Hochul’s decision as an example of a leader being more responsive to her duty to constituents than to “earning a reputation for being business friendly.”

Norton reminded the audience of the time when Uber, developing its own robocars, was looking for states to test its AVs. Uber first tried California. But as Norton explained, “California said, ‘Our people are not your experimental test subjects, at which point the governor of Arizona raised his hand on Twitter and said, ‘We have open roads and open minds come here. Months after that, Elaine Herzberg was dead.”

Norton’s reference was to an Uber self-driving test vehicle that struck and killed 49-year-old Herzberg on March 18, 2018 in Tempe, Arizona.

The power of policy makers and regulators

Given a chance to offer advice to the AV industry, Zipper made clear, “It wouldn’t be to Waymo. It would be to Congress … It would really be primarily to National Highway Traffic Safety Administration and to the Federal Trade Commission, who have, in my view, just completely abdicated their obligation to think ahead, on behalf of all of us and on behalf of the industry, to protect us from some bad outcomes in the future.”

He added, “The real failure here is on the regulators.”

Perhaps the most significant element in this controversy is the amount of money the AV industry has spent lobbying public officials in a position to decide transportation policies.

Zipper said, bluntly, “Everyone who says they’re a public affairs person for an AV company is a lobbyist. That’s just a factual statement.”

He cited a city councilor in Washington, D.C. who disclosed to him that every single member of the council in his district is “on retainer with Waymo.”

The balance of power, Zipper concluded, is “tilted so far on the side of the AV industry” that no real dialog is happening.

Clearly, the panel went after the most fundamental AV question of all time: What is AVs’ raison d'être?

Or more precisely, what problems do we want AVs to solve?

As Sam Schwartz and Kelly McGuinness reminded everyone in their Street blog piece, “Cities that don’t set the rules upfront end up living by someone else’s.”

Related stories:

With Waymo Testing Halted, We Have A Rare Chance to Get Ahead of the ‘Driverless Revoluion'

By Sam Schwartz and Kelly McGuinness (Street Blog)

Safety Is Not Their First Priority

Nor should it be. “Safety first” is the wrong way to message prioritizing autonomous vehicle safety. There is a better way.

By Phil Koopman (Phil & Junko on AV Safety on Substack)

For a thorough understanding of the massive changes to the automotive industry in recent history, there’s no better source than Riches, vice president of automotive at TechInsights. Affiliated with different market research firms over the years, Riches has spent his entire professional life as a key witness to the global auto industry.

When I called about his 30th anniversary last month, my first question was: “You aren’t going to retire, are you?” He responded, “Are you kidding? This industry has got so much going on. It’s too interesting to retire.”

Blame it on Apple

In the last 15 years, automakers have been forced, more and more, to alter the way they do their business. Riches blames this on Apple.

He said, “Of course, I’m massively oversimplifying.” But his emphasis is that the auto companies have begun to recognize that forces outside their own industry can no longer be ignored.

Take, for example, the mobile phone industry’s massively accelerated development cycle for new products and technologies.

This compressed process of deploying new features and functions has prompted the automotive industry to “suddenly march to someone else’s timeframe,” Riches observed. Auto-centric apps like navigation turn out to be “completely out of step with traditional automotive industry timescales,” he added. In short, “Automakers had never imagined that their business would be affected by a totally different industry like smartphones.”

Not invented here

This turmoil was also triggered partly by automakers’ affliction with the “Not Invented Here” syndrome. Consider, for example, how long it has taken for automakers to adopt Ethernet

Automakers’ persistent desire to reinvent and reinterpret innovations already broadly accepted elsewhere has dogged the industry.

Riches said: “What was designed to keep the auto industry in control never worked.” Hence, many technologies, notably connectivity and interface, are largely defined by Google and Apple.

Of course, in their defense, the automakers’ mantra can be summed up as, “It’s an automotive thing. You don’t get it.” This presumption, which Riches has seen again and again, is precisely the culprit. As he explained, “They look at a technology elsewhere and they think, yeah, that’s 90-percent perfect. But we need to tweak it, because they don’t really understand us.”

Tesla’s Riches

Automakers’ bias — thinking that they’re special — has also sometimes blinded them to the big picture, and to new trends emerging outside their clubhouse.

Case in point: Tesla.

“A lot of automakers, probably myself included, actually missed what I now see made Tesla so special and unusual,” noted Riches, After all, industry observers studying the industry’s history were aware that anyone who tried to start an auto company pretty much failed, until Elon Musk.

Tesla founded a niche as an EV company, but Riches does not think that’s the whole story. Tesla beat the odds as “the first ‘holistic company … when much of the auto industry was, and still is, so deeply siloed.”

Similarly, automakers’ decades of experiences developing individual domain-specific features are getting in the way of cross-domain innovations, observed Riches. Fresh challenges facing automakers include “combining compute and sensing inputs and actuation across all of those different areas,” he said. “If you couldn’t do that effectively at scale and at speed, you’d find yourself at a massive competitive disadvantage.”

Here’s my conversation with Ian Riches:

The videopodcast is available on my YouTube channel: Junko’s Talk to Us.

For clickable content links, please see the YouTube show notes.

Chapters:

(00:00) Intro

(01:47) How do you become an analyst?

(04:24) More disruptions affected the auto industry in the last 15 years

(06:57) Automakers have suffered from the Not-Invented-Here syndrome.

(10:30) Which technologies failed?

(15:37) What made Tesla so special and unique?

(21:51) Throwing more software engineers isn’t the way to solve the challenges

(24:55) Bifurcation of the auto industry

(28:33) How do you build your own brand?

As I noted in the inaugural “Phil & Junko on AV Safety” video podcast, I’m glad that we’ve reached the point where the issue of human remote assistance operators is no longer shrouded in a mystery. Now, it’s a topic of investigation, led by Senator Ed Markey (D-Mass).

Once reluctant to acknowledge the existence of human remote operators and their reliance on them, robotaxi companies now have to walk back their narrative.

While AV companies stick to their claim that autonomous vehicles are smart enough to make safety-critical driving decisions by themselves, the truth is out there: Driverless cars are not human-less.

Seeking to get to the bottom of remote assistance operations, Sen. Markey asked AV companies 14 questions.

He asked, for example, the roles and responsibilities of the remote assistance operator. He asked if any remote operators are located outside the United States and wanted to know the average and worst-case latency between a vehicle and its remote operator. He asked how often remote assistance sessions are invoked. He sought descriptions of the training and qualification of such remote operators and he requested a copy of any internal policies or standards that govern remote assistance operations.

Markey’s questions appear to cover a lot of ground. But let’s not be naïve.

When matters turn political, lawmakers (and the media) tend to simplify, framing issues in headline-length phrases that stir up social, political and economic controversy—i.e. which passport do remote assistance operators carry, or where did they get their driver’s licenses?

The danger in this approach is that it can overlook a lot of nuance inherent to the different roles played by remote operators. Without attention to this sort of detail, AV companies can find ways to sidestep the responsibility to keep their remote operators accountable.

In our podcast, I asked Phil whether Sen. Markey asked AV companies the right questions. Did he illuminate the real issues of remote operations? What did he miss? Where does his investigation need to focus?

Phil believes that Sen. Markey (and his staffers) were asking some of the right questions, but missed a few important points.

The issue in point is accountability.

Sen Markey asked questions about the “qualifications” of human remote operators. But he fell short of clarifying those operators’ accountability.

“The question is not what driver’s license those remote operators have,” stressed Phil. “ The question is whether they are qualified to operate” remotely the type of vehicles for which they are responsible.

Further, Phil pointed out, the source of a remote operator’s license is less important than where the operator is physically located.

“This has to do with law enforcement,” he explained. Phil posed the hypothesis of a remote operator who shows up for work, drunk, impaired, or sleep deprived… he is just not fit to drive.” When a mishap occurs, because this impaired operator made a bad call, is there a traffic cop available. Not if the car’s in San Francisco and the operator is in the Philippines, or even in just a different state.

Phil asked, “How does the State Police administer a sobriety test or detain or question a person who is three time zones away? Is there an interstate agreement that the other state police will come in as proxies?”

Right now, these questions are entirely rhetorical.

Enforcement can be arranged. But, as Phil reiterated, in the absence of a regime that monitors mishap and response, “how do you have accountability for a remote operator who’s driving impaired? How do you even do that?”

Ep 1 of “Phil & Junko on AV Safety” is now available on YouTube.

Watch and you’ll learn the details of how we define the six roles of remote assistants. We explain why these roles are relevant to safety. We discuss Senator Markey’s investigation and explain the importance of location, time lag and other nuances of remote robocar assistance.

Subscribe for free to receive new posts and support “Phil & Junko on AV Safety.” While I continue with “Junko’s Tech Probe,” Phil Koopman and I weigh in on AV safey matters in the new substack below.

Phil & Junko on AV Safety

Related story:

One of the most under-discussed challenges facing automated vehicles is not technological. It is institutional. What is missing is a serious public debate about how, where, and under what conditions these systems should operate, and who is responsible when they fail.

This is not just a deployment challenge. It is a familiar governance problem: institutions often lag behind the technologies they are meant to manage. Companies build products around market incentives, while public institutions are left to adapt after the fact.

Companies deploying robotaxis are expanding aggressively across major U.S. cities. Cities not yet selected for deployment increasingly worry they are being left behind in what investors and technologists frame as a defining shift in how people live and move.

But that assumption deserves far more scrutiny than it receives.

In 2022, a driverless vehicle in San Francisco blocked a fire truck responding to an emergency. Incidents like this are still treated as edge cases, exceptions rather than signals, yet they point to a deeper issue: how these systems behave in complex real-world environments shaped by human authority.

News reports show city officials and first responders are increasingly dealing with decisions made remotely, often with limited visibility into how these systems function or fail. When things go wrong, whether a vehicle blocks an emergency responder, an unexpected maneuver creates a traffic jam, or a pedestrian is struck, the lines of accountability begin to blur. Fault is distributed across corporate structures, software versions, remote operations, infrastructure, and the human behavior that remains part of the system. That diffusion makes ordinary accountability tools, including traffic citations, much harder to employ without regulatory change.

Changes in the math

A behavior deemed acceptable in a pilot deployment becomes problematic when repeated over thousands of miles each day. The math changes, and so does the nature of risk. The real questions are: Who is responsible? What level of systemic risk is acceptable? And, who gets to decide?

The technology is advancing. That much is undeniable. But the ability to deploy is not the same as readiness to operate at scale. This distinction is being glossed over by elected officials, by industry, and by the broader assumption that technological progress can outpace governance without consequences.

Communities choosing not to rush deployments may be making the more rational choice. Cities that position themselves at the forefront of deployment are also choosing to serve as live test environments for some of the most complex systems ever introduced into public space. Their residents, in effect, become unwitting participants in a real-world experiment, one that is consequential, imperfectly understood, and often opaque to oversight. Automated vehicles are an example of a broader pattern: society often grasps the full implications of new systems only after they are embedded in everyday life. We should not have to rely on courts after the fact to restore balance to systems that were poorly governed from the start.

On most days, the system works. Riders experience convenience and efficiency. These are real signs of progress.

But not every day is routine. Scale does not just expand performance. It amplifies failure.

When systems fail or behave unexpectedly, the consequences compound, affecting traffic flow, mobility, and public trust. In these moments, accountability does not just blur. It fragments. Operators blame an edge case. Local officials discover they have limited recourse, and the public bears the cost.

This is not an argument against automation. The future of automated mobility is real, and the potential benefits are significant. The path forward is unlikely to be the clean, rapid transformation that investors and technologists prefer to project. It will be uneven, iterative, and constrained as much by policy and public acceptance as by technological progress.

Questions that can no longer be put off

These gaps raise questions that can no longer be put off:

What level of data transparency is required to assign fault and ensure accountability? Today, the data generated by these systems remains largely inaccessible to oversight authorities, even as vehicles operate in public spaces. Cities can demand dashcam footage from Uber drivers but cannot access AV telemetry.

How should these systems be taxed? Unlike traditional taxis, these systems operate within regulatory and infrastructure frameworks that were not designed for them. That raises a basic question: who ultimately bears the cost of deployment?

What obligations do operators have to insure against failures that disrupt emergency response, accessibility, and mobility during critical events, not just crashes?

How should policymakers think about workforce impacts, including not just job loss but where those losses fall and who bears them?

And perhaps most fundamentally: how should safety be defined?

“Safer than a human” is often invoked as a benchmark. That framework is incomplete, not because it is wrong, but because it is not enough. Safety is not static, nor is it purely statistical. A system that reduces certain risks while introducing new, less understood ones may struggle to earn public trust. What is acceptable today may not be acceptable tomorrow, particularly if these services worsen congestion, reduce public transportation use, and undermine sustainable mobility goals.

These are not questions to answer after the fact. They are foundational to whether automated vehicles serve the public interest or become another technology optimized for private gain at public expense.

Challenges for policymakers

The challenge for policymakers is not whether to support innovation, but how to shape it. History suggests that when deployment outpaces governance, the public pays the price of rebuilding trust. And rebuilding trust takes years.

The timing of this conversation matters. Public discourse rarely keeps pace with deployment at scale. It must evolve alongside it.

Automated vehicles are not arriving all at once. They are emerging through a long, uneven evolution shaped by technology, human behavior, economics, and policy. Technology and economics move fast. Governance moves slowly and often reactively. This mismatch creates pressure to deploy before the rules are written.

That reality should push city and state leaders to move faster on policy and regulation. Perhaps not at the pace financial markets or technologists would prefer, but government must move faster now.

There is no requirement to be first. Being first carries costs in the form of operational disruptions, liability, and infrastructure strain. The benefits, meanwhile, accrue more broadly once the systems mature. Early adoption attracts venture capital and headlines. It also brings lawsuits, congestion problems, and political blowback.

The more strategic choice is often not to lead the experiment, but to observe it and write the rules while others prove the technology works.

As policymakers grapple with the future of transportation, it is worth remembering that slowing a technology’s rollout can help ensure public benefit keeps pace with corporate incentives. It also gives government time to build the governance these systems need to succeed.

In the case of automated vehicles, restraint is not resistance to innovation. It is how innovation succeeds.

Getting this right is not about moving first. It is about building systems the public is willing to trust.

— Guest columnist Bryan Reimer is a research scientist in the MIT Center for Transportation and Logistics, and a researcher in the AgeLab. This piece is co-posted by Reimer on LinkedIn.

By now, you might have heard through the techvine that I’m joining Phil Koopman on a new project called “Phil & Junko on AV Safety.”

We embark on this endeavor to focus attention on the safety issues of autonomous vehicles, and more broadly, embodied AI systems.

Of course, I’ll continue posting news & analysis on business and technology matters, along with market/industry observations at “Junko’s Tech Probe.” Anyone who has followed me all the way back to my days as editor-in-chief of EE Times knows well my interest in the territory where technology meets people. My coverage ranges from semiconductors and embedded systems to automotive and Artificial Intelligence.

Naturally, safety is a key topic in the tech-meets-people realm.

However, lately, I’m increasingly frustrated. I’ve realized that chasing one safety incident after another in my reporting doesn’t exactly take my readers where we need to be to understand the fundamentals of safety engineering.

Phil has long enriched my understanding as a sounding board. We’ve come to know each other better, and more importantly, to enjoy conversations both technical and personal.

I hope that our new Substack “Phil & Junko on AV Safety” — where we collaborate on video podcasts and develop story ideas — will become a useful resource for you. We’ll be exploring topics that seem broadly understood but worthy of re-examination. We’ll dig around in areas where most of us don’t know much, or hardly anything at all.

If you’re even remotely curious about safety matters, we’re inviting you to subscribe free to “Phil & Junko on AV Safety”. Sign up now, so you can receive our inaugural email newsletter, and you won’t miss our upcoming podcast (Ep 1, this Saturday!).

Phil & Junko on AV Safety

Phil Koopman & Junko Yoshida weigh in on Autonomous Vehicle Safety and related topics

Below is a snippet of our newsletter that just went out today:

Weekly News: Phil & Junko Launch Our Newsletter

By Junko Yoshida & Phil Koopman

Welcome! This is the launch issue of what we plan as a mostly-weekly newsletter. It will contain brief summaries of other materials we’re creating, including thought pieces and video podcasts. We’ll also provide curated pointers to things we think everyone should be aware of in the Autonomous Vehicle (AV) safety space.

Greetings from Junko Yoshida:

The “Phil & Junko on AV Safety” is a collection of podcasts, thought pieces, news analysis and opinions in which I get to collaborate with Phil Koopman. This is a dream-come-true moment for a longtime technology/business reporter. Before becoming an independent journalist, I was editor-in-chief of EE Times covering semiconductors, embedded systems and software, mobile phones, and automotive.

Reporters chase news and write our stories as events unfold, working to synthesize what we’ve learned into analysis.

But when the world is moving so fast and technologies growing increasingly complex, I worry if we are serving as mere stenographers of AV companies’ claims.

While tech companies’ investment and partnership news are vital to measure their business prospects, we need to dig deeper in coverage of AV safety. The demands of the AV safety beat require close scrutiny on the technology itself (not just about TOPS of an AV’s brain CPU). It’s critical to understand the roles played by regulators and the legal tangles victims are bound to face when AV technology goes haywire.

As a reporter, I wonder. Are we asking the right questions? Perhaps, more importantly, how do we know we’re not being taken for suckers?

The best insurance for me—and for readers—to avert this danger is this collaboration with Phil, who is recognized as one of the foremost experts on AV and embodied AI safety. Our mission is to probe this vital issue not as an industrial “race” with its winners and losers, but as an unfolding chronicle of how to define and share, without hype or naiveté, the benefits of AV safety and the pitfalls of AV hubris.

Greetings from Phil Koopman:

I’m excited to be starting this journey of education and safety advocacy with my friend Junko. We’ve had so many talks over the years about various automotive safety stories, but this is the first chance we’ve had to work with as co-authors and collaborators. She brings impressive journalistic chops to the collaboration, and I’m delighted to have the chance to work with her.

For those who don’t know me, I’ve been working on self-driving car safety and more general embedded system safety for about 30 years as a professor at Carnegie Mellon University. I recently switched to an emeritus role there, which amounts to an inactive status. (“Retirement” is perhaps too strong a word.) That gives me more time for activities such as consulting and writing for this collaboration. My professional background is computer engineering, embedded computer systems, and system safety. That includes time as a US Navy submarine officer, a CPU designer, an embedded system engineer, and running hundreds of design reviews for industrial embedded system projects across numerous industries.

… To read the rest of the first “Phi & Junko on AV Safety” newsletter, please sign up.

Phil & Junko on AV Safety

Phil Koopman & Junko Yoshida weigh in on Autonomous Vehicle Safety and related topics

Sony-Honda Afeela press conference in Jan., 2024 at CES (Photo Credit: David Benjamin)

Sony Honda Mobility, the joint venture between Japanese electronics ‌giant Sony and automaker Honda, announced Wednesday that they would be scrapping its Afeela electric vehicles project.

The JV declared Afeela “unviable” after Honda scaled back its EV plans. In turn, Sony, a non-automaker, chose to drop its own EV ambitions.

I hesitate to say I told you so. But this retreat didn’t surprise me.

Any reporter who has attended enough press conferences sometimes gets the feeling that this latest Next Big Thing is “doomed,” — whether the thing is a joint venture, a new technology or the prototype of a brand-new can’t-miss product.

Usually, the “tell” shows at a press conference that’s largely—ominously—news-free. The spiel offers too much marketing, too much corporate branding and a lot of posturing. At CES 2024, the unveiling of Afeela — the Sony-Honda JV— was exactly that.

Sony, of course, is a respected Japanese consumer electronics giant, and Honda an equally renowned Japanese carmaker.

But beyond their notoriety, past glories and the esteem of their Japanese countrymen, what did they bring? What could two legacy companies from different industries possibly do to surprise an automotive market already disrupted by new entrants (Chinese OEMs), new technologies (electrification, software defined vehicles and AI) and supply chains that have become increasingly complex and geographically divided?

Despite their high hopes for Afeela, neither company had answers for these questions, nor did they seem to have even asked them. What they learned— if they did—is that in both of their industries, hubris is the great leveler.

Below is a story written and filed by our colleague David Benjamin right after the Afeela press conference at CES 2024.

CES ’24: ‘I feel ya, man.’

By David Benjamin

“… Feelings, whoa, oh, oh, feelings/ Whoa, oh, oh, feel it again in my arms/ Feelings, whoa, oh, oh, feelings/ Whoa, oh, oh, feel it/ Whoa, oh, again”

— Morris Albert

LAS VEGAS—Watching a relay team of Sony, Honda and Microsoft execs at a Consumer Electronics Show (CES) “press conference” Monday as they meandered toward the unveiling of a prototype “software-defined vehicle” (car) called Afeela, I found myself idly pondering the possible etymology of its name.

I suspected that “Afeela” might be a corporate appropriation of the bonding device, likely derived from hiphop culture: “I feel ya.” Applying this coinage, it’s easy to discern a motif that recurs annually and persistently at CES.

The masters of the consumer electronics gadget industry say that they want—above all other considerations—to usher “humankind” into “experiences” that will make us all feel better. Really! Why else go into business?

The ideal CES forum for conveying this altruistic theme is either a “keynote” address—the Consumer Technology Association stages a dozen or so every year—or a “press conference.”

You might have noted the quotation marks. This is because the so-called “press” at CES is a motley melange of bloggers, vloggers, influencers, fanziners and plain old fans, early adapters, shills, hackers and rubberneckers. There remain a few old-school ink-stained wretches like me. But we’re a dying breed.

Demographics aside, I’ve never endured a CES “press conference” during which any conferring occurred. The invited media are offered no opportunity to query the corporate leaders on stage. No reporter compelled—if he or she wants a seat—to queue up for one of these hypefests has ever tried, or expected, to ask a question. It’s all pedigreed dog and thoroughbred pony.

The consolation in this one-way news “conference” is that it’s largely news-free. For example, the star gadget in Sony’s show, the Afeela-mobile, was at least six months old. I looked up comments made in June, about the car’s prototype, by Nakul Duggal, who’s head honcho for automotive stuff and cloud computing at Qualcomm, the company responsible for Afeela’s in-car electronics hardware.

Duggal proved an excellent pitchman for the touchy-feely angle on gadgetry. He referred to the inside of the car as an “ecosystem,” thus suggesting the harmony of all living things—animals, trees, butterflies, people, cars, chainsaws. He called the Afeela interior a “digital living space,” echoing other techno-promoters who foresee a car—twinkling and thumping with video screens, dashboard game displays, rock ’n’ roll and holographic porn stars—as a “digital living room.”

I’ve suggested to a consumer electronics expert that, nowadays, the only people who “live” in their cars do so involuntarily. They’re homeless. But I’m reminded that these people are residents of old-fashioned, internal-combustion “analog” living spaces that will soon be as outdated as … well, me.

CES press conferences typically present two sorts of products. The first is something that’s been on the market for a while but it’s been “reinvented” to keep it up to date or to correct flaws that threaten its survival. The other, like Afeela, is a techno-miracle, still on the drawing board or spinning ’round the test track, that will—when it’s ready—transform society and gently guide people toward a utopia for which the current crop of humans might be ill-prepared to appreciate.

Or, as Marty told his parents’ generation in Back to the Future, “I guess you guys aren’t ready for that yet. But your kids are gonna love it.”

In thirty-odd years at CES, I’ve witnessed the coming and going of many imminent miracles. I’ve seen more artist’s renderings of Paradise—projected on twenty-foot screens eighty feet wide in sprawling ballrooms at the Sands, Venetian, Mandalay Bay, etc.—than I could count. For instance, I’m still awaiting the proliferation of intuitive smart homes that know where you’re going, when you’re coming back, what you want from your smart fridge after you’ve crossed the smart threshold and whether you need to go to the smart toilet for number one or number two. These smart homes will fill smart cities whose smart streets teem—safely—with driverless cars and trucks, buses, trams and dumb kids on smart bicycles.

“Look, ma! No hands!”

I cautiously expect some variation of this Jetsonworld to emerge eventually, although piecemeal and subject to the economic straits of its inhabitants. It will look nothing like the giant slides—augmented with strobes and pyrotechnics—in the Palazzo ballroom. I hope to Christ it doesn’t resemble Blade Runner.

But the future’s not the point.

The mission of CES, every year, is not to mark tangible progress toward a time in which the advance of technology fulfills the promises made by marcom sharks. Like any trade show, CES is a pep rally. CTA chief Gary Shapiro’s job is to create—on every exhibitor’s behalf—a buzz so loud it drowns out the outside world. CES returns bigger and better every year so that all those credulous influencers will gape and go “ooh” and take photos, make videos, flood TikTok, Instagram, YouTube, etc., with images of the Next Big Things, thus allaying the anxieties of the investment community and keeping the wolves from the door.

CES also has a psychological role, to stir the emotions of participants about its products and to assure them that they’re all making gadgets that are worthwhile and doing business that’s good, proper and virtuous. These are mothers of invention who are NOT only in it for the money.

But listen carefully. Sony’s presenters assiduously referred to the employees making their films, music and games as “creators,” shining on them a weird glow of godliness. By and by, however, the script slipped. All those songs, movies, screenplays and medieval fantasy worlds fell back under the heading of “content.” This candid slip of the tongue restored Sony’s artists, writers, lyricists, singers, composers, actors, directors and animators to a familiar, expendable commodity: “content providers.”

Finally, what about us fragile human creatures whose feelings—not our wallets—are the reason for all “innovations” to exist, progress and create new things that make the previous things old and useless. They want for us, “the joy of experiencing life’s potentialities, “a new experience that defies conventional wisdom,” “a creative entertainment space in an immersive experience.”

Eventually, as each presenter drones on about the human “experience,” he or she reverts to words like “customer” and “consumer,” which after all gives away the game—the “C” in CES.

But never mind ulterior motives. After the Sony people finally pushed the world’s only Afeela onto the stage, I took a quick photo of the car and headed toward the exit. Except I couldn’t.

The aisles were filled. I turned away from the stage to a wall of faces, camera lenses and uplifted smartphones that illuminated eyes round with wonder. They were packed like poultry in a factory farm, credulous and awestruck, gaping at a car that isn’t yet a car, peering into a future that’s already six months old, willing to believe a promise that lies somewhere between a pipe dream and a crap shoot.

Which is probably why CES takes place in Vegas, where pilgrims come to make their fortune and almost everybody goes home busted. Some even end up living in their cars.

—Editor’s note: The story above by David Benjamin was initially publisheed in January, 2024 on The Ojo-Yoshida Report.

ST64UWB solution extends range and enables best-in-class hands-free experience. (Image: STMicroelectronics)

Ultra-wideband (UWB), a short-range wireless communication technology, is hardly new.

The technology going back to 1880’s was originally invented and later improved to send radio transmissions across the Atlantic Ocean. Later, impulse-based wideband radio technology was developed to determine the range, angle, and velocity of objects mainly for military applications such as location finding and communication technology.

In 2002, the Federal Communications Commission allowed certain types of UWB devices to use spectrum occupied by existing radio services. That decision launched a slow-motion race to commercialize UWB for civilian uses.

In 2019, Apple implemented UWB in consumer electronics, embedding it in its iPhone 11. Since then, UWB has shown up in devices that include wearables, automotive, tags and smartphones. But UWB, widely seen as a chronic underdog in radio technology, has not exactly taken off.

I bring up UWB because STMicroelectronics this month came to the Embedded World show in Nuremberg, Germany to launch its home-grown ultra-wideband (UWB) chip family, dubbed ST64UWB.

Notable about ST’s announcement is ST64UWB is an industry-first system-on-chip supporting the latest ultra-wideband specification, IEEE 802.15.4ab.

This announcement leapfrogs NXP Semiconductors, which has led the UWB market. NXP’s launch of an 802.15.4ab-compliant chip is due later this year.

ST, UWB and Automotive

Betting big on 802.15.4ab, ST professes that it can alter the trajectory of UWB adoption. ST also appears focused on automotive applications as its foothold in the UWB market.

ST describes the standard, built on existing IEEE 802.15.4z UWB wireless technology, as “a big leap in reliability, range, and radar, effectively eliminating ‘blind spots’ in challenging multipath environments.”

(Source: STMicroelectronics)

Enhancements in UWB performance were enabled by “multi-millisecond ranging (MMS) and narrowband assistance (NBA),” explained Rias Al-Kadi, general manager of the ranging and connectivity division at STMicroelectronics.

Most notably, the new standard can fix what Al-Kadi considers “the most annoying problems with the wireless car key built on the existing UWB standard.” When the key fob is in a back pocket, it won’t unlock a car because the signal is blocked by the car owner’s body.

ST is also hoping that radar mode enhancements made by the IEEE 802.15.4ab will enable further uses of UWB inside vehicles. The new UWB standard can literally improve use cases such as detection of children or pets left behind inside vehicles. This is a potentially life-saving feature recommended by Euro-NCAP, explained ST.

Beyond using UWB as a camera-free, wireless occupancy sensor, ST’s Al-Kadi noted that the new standard extends the operating range to detect life forms on the road. In autonomous vehicles, for example, he said, UWB can be a safety tool for perceiving a pedestrian long before a camera can do so.

Millions of applications

Among the virtues of UWB technology is that it offers a very low energy level for short-range, high-bandwidth communications over a large swath of radio spectrum. By offering secure ranging and precision sensing abilities, UWB can provide fast and reliable data transfer.

Despite such advantages, UWB’s broader technology adoption has lagged for a number of reasons.

First and foremost, UWB comes with a higher chipset price compared to other popular wireless technologies. According to Omdia, a basic UWB IC at volume would cost about $4, ten times more than a secure Bluetooth Low Energy IC.

Another issue is the limited number of semiconductor companies playing in the UWB chip market. ST is only the fourth major company to commit to a UWB market thus far dominated by Apple, Qorvo, and NXP.

Perhaps most important is that UWB faces an embarrassment of riches in the potential applications of radio technology.

UWB applications for consumers range from tagging devices for finding their location in real-time to easing wireless file transfers and turning smartphones into car keys. For industrial applications, UWB’s effective proximity sensing enables safer operation of tools and machinery by equipping workers with UWB-enabled devices or badges. UWB also makes possible more precise and efficient asset management in industrial operations.

“There is a lot of stuff you can do with UWB,” said Edward Wilford, senior research director for automotive at Omdia. However, crucially, “the scatter-gun approach—to show a million things you can do with UWB—isn’t as helpful as finding one killer app.”

(Source: Omdia)

How big a deal will the new UWB standard be?

Omdia’s Wilford is unsure how big the new UWB standard will be. He said, “I think the capabilities for radar-like behavior are interesting, and more important than the extended range for automotive applications. The demonstration at ST showed promise, but I don’t yet see the real-world use cases for all the features.”

Consider the example of driver and occupant monitoring, he said.

“There is already a legal requirement in many countries (or will be soon) to have a camera monitoring the driver for attentiveness, signs of drowsiness, etc. If you have a camera, do you need to have a radar system based on UWB? This is why a lot of people have abandoned other ranging tech, such as WiFi radar, I suspect … There is usually already a working technology in place. So while I’m bullish on UWB and happy to see ST in the mix with a great product right out of the gate, I don’t know that it upends the market.”

Perception, cost and ecosystem

Asked what he believes has caused UWB to stall, Wilford said, “I think a lot of it is perception, cost and ecosystem.”

He explained, “UWB is still seen as the AirTag tech (even though BLE does 99% of the work). Its security features aren’t as well known, as it relies on a different form of security rather than traditional cyphers and keys.”

Noted earlier, the cost difference between UWB IC and BLE remains large. For UWB applications, a car would need as many as six chips to cover all the approaches, plus two key fobs, said Wilford.

That adds up. It projects UWB as a volume market for chip suppliers but not so much for its OEMs buyers.

One of the strengths of UWB is that wireless technology is heading for a variety of devices. In theory, a combination of different devices already embedded with UWB poses intriguing applications. But it also creates a huge dependency on its ecosystem. “You need a car and a phone and a smart home set-up (and also maybe a watch or wearable) with UWB to really make the ecosystem work,” observed Omdia’s Wilford. “Getting to that critical mass isn’t easy.”

UWB’s ecosystem play

Naturally, UWB’s applications aren’t limited to automotive. A UWB ecosystem can play a role in the smart home of the future, Wilford offered this example:

I should be able to put my phone in my pocket, walk out of my house. The door locks behind me, car unlocks in front of me … I never have to take this [UWB-featured device] out, or my watch as long as it’s on my wrist and unlocked.

Those are the nice features that I think would be enough of an ecosystem reinforcement.

There are also safety features at home, which are enabled by UWB.

NXP showed a gas range at home that would only turn on if an adult with a phone was nearby. So, you’re upstairs and your child’s downstairs and they can’t turn on the stove, no matter what they do.

NXP’s views

Bernhard Grosswindhager, product market manager for UWB at NXP, acknowledged the increased performance of the new IEEE 802.15 ab. He said that more robust UWB communication links can lead to more reliable UWB-based systems, including hands-free automotive and building access.

“As a key contributor to the new IEEE 802.15.4ab, we are working on the next generation of UWB supporting 802.15.4ab and expect to make an announcement on this later this year,” he said. “We believe we will be the only vendor globally to support products in automotive, industrial/IoT and mobile applications.”

However, he pointed out, “It’s worthwhile to mention that a lot of OEMs are still planning to rollout 802.15.4z-based systems. Therefore, we expect the rollout of 802.18.4ab-based systems will start slowly and will gradually replace 802.15.4z-based systems.”

Bottom line

The potential for UWB seems to be big. Yet the big barrier for broader adoption is its cost. Omdia’s Wilford believes UWB needs to get the price down to $2 a chip. “Then we see OEMs start to include UWB as a default,” he said.

The other option, he added, “is to get the insurers or regulators to force their hand. OEMs won’t mind if everyone gets extra cost per vehicle, but they don’t want to be first movers, especially at the mid to low end where margins are so thin.”

(Image: iStock)

Two German automotive giants, first Mercedes-Benz, then BMW, confirmed that they are dropping SAE Level 3 vehicles from their product portfolios.

This comes as no surprise to industry observers like Ian Riches, vice president of automotive practice at TechInsights, who has been flagging the machine-human handover problem since the breech birth of L3.

Judging from their statements, Mercedes-Benz and BMW appear to frame the failures of their L3 vehicles because they were just too expensive. “Demand for Level 3 wasn’t at a profitable level,” as BMW tells us.

The way I see it, that’s not all.

I frame the issue as a matter of trust. Car companies have failed to build trust among the sort of consumers who would be receptive to highly automated vehicles.

It’s worth remembering Mercedes-Benz’s boast when it launched L3 vehicles in January 2023:

While active, DRIVE PILOT unlocks activities on the central display so the driver can play games, watch videos or take advantage of in-car entertainment features.

Capable of detecting your surroundings, conditionally automated vehicles can make informed decisions for themselves.

Although it did warn that its L3s “require the driver to remain alert and take control when requested,” Mercedes-Benz never demonstrated the intellectual honesty to acknowledge the obvious contradictions in the company’s own statement.

Benz offered no clear guidance about what its L3s can and cannot do and how quickly a human drivers might tune in to the fact that the car can’t “make informed decisions on its own.” Moreover, the carmaker neglected to mention who gets the blame if, for example, the vehicle failed to inform a human to hey, quick! Take the wheel!

Yes, ’twas price killed the L3. But don’t underestimate the power of consumer intelligence.

Consumers saw through the manufacturers’ bait-and-switch pitch. They said, “L3 can give you back time.” But your time was up when “vehicles require you to remain alert and take control.”

Perhaps most important, according to Colin Barnden, principal analyst at Semicast Research, “Consumers can do a cost/benefit analysis and work out L3 doesn’t offer an attractive value proposition.”

A pathway to L4

While most automakers today seem ready to put the L3 fiasco in the rearview mirror, a trend visible through the windshield is auto and technology companies using a brand-new technology to sell new L2+ or L2++ vehicles. They are promising that the new platform will blaze the trail, someday soon, to a fully autonomous L4 vehicle.

No kidding.

Consider Nvidia GTC, a global artificial intelligence conference this week in San Jose. The show was light on automotive-related news, leaning instead on Nvidia’s Physical AI pitch. But CEO Jensen Huang announced that BYD, Geely, Isuzu and Nissan are all building level 4-ready vehicles on the NVIDIA DRIVE Hyperion platform.

Similarly, earlier this month, Qualcomm announced a technology partnership with London-based AV company Wayve to integrate Wayve’s end-to-end AI Driver software with Qualcomm’s Snapdragon Ride platform. Their promise? To provide “a production-ready ADAS/AD system ranging from hands-off driver assistance to eyes-off automated driving, with a path to Level 4 robotaxi applications.”

As TechInsights’ Riches noted, these announcements show that “things ARE happening” on the auto industry’s L4 ambitions. But he also projected a timeline for volume L4 that “still looks well into the 2030s.”

How is this different from Tesla’s playbook?

The latest roadmap from L2+, L2++ to L4 might intrigue the early adaptor clique, but how does it restore trust among ordinary consumers? How difficult will it be, for both human and vehicle, when the car vehicle upgrades from L2++ to L4?

Will Nvidia’s Hyperion platform, together with Alpamayo, smooth the transition?

Will Wayve’s ground-breaking E2E platform succeed in magically solving the pesky real-world problems that have frustrated robotaxi operators?

It’s worth noting that this new pitch — drawing a straight line from L2++ to L4 — is not so different from Tesla’s FSD playbook.

Just this week, the National Highway Traffic Safety Administration’s Office of Defects Investigation (ODI) upgraded its initial probe into Tesla's Full Self-Driving (FSD) to “engineering analysis,” its highest level of scrutiny. According to NHTSA, “The focus of this investigation will be to assess the system’s ability, when encountering reduced roadway visibility conditions, to detect degradation and alert the driver with sufficient time to respond.”

Phil Koopman, safety expert, professor emeritus at Carnegie Mellon University wrote on his LinkedIn post:

My take: if any system that provides sustained control of steering lets itself operate outside its designed environment (ODD) without warning the driver, a rebuttable presumption of responsibility for any crash should be on the manufacturer, not the driver. Arguing that the driver is responsible for enforcing an ODD limit not readily perceived by human senses (e.g., exactly how much haze is “too much”) puts that person in an untenable position.

ODD constraints

ODD constraints won’t suddenly disappear when SAE Level transitions to higher levels.

The main reason L3 flopped is its ODD.

TechInsights’ Riches observed, “The tight ODD constraints [in L3] weren’t arbitrary—they were a direct consequence of the handover problem.”

The use case for BMW’s “Personal Pilot L3” was navigating traffic jams or slow-moving traffic.

Clearly, “L3 puts the driver in an awkward position: relieved of the driving task but required to be ready to resume it at short notice.” That, in his opinion, “made the system almost irrelevant to daily life. If it was €600, fine. €6,000? No chance,” said Riches.

But that’s not all.

Riches touched on the problems of consumer trust. It’s not clear he said, “whether many consumers trusted the technology enough to use it even if it had been cheaper.”

The ODD for L3 was so narrow that it afforded little chance to build confidence in it. He said, “All the user research I’ve seen shows that trust is weakest in the older demographics.” BMW’s L3 was offered only for 7-Series.

It’s already clear L4 won’t solve all the problems posed by L3. L4 robotaxis have already faced basic—and unforeseen—AV problems.

Semicast Research Barnden believes, “There is no quantity of sensors or compute that can ensure a machine driver performs in all scenarios in a complex dynamic system, of which public roads is an example.”

The heart of the issue reveals itself when things go wrong. “For a human driver, we blame the human. They made a mistake, but we can reassure ourselves that it was a one-off. That particular human made an error,” said Barnden. But a person can learn the lesson and do better next time.

Not so much for a machine driver.

To be clear, Waymo does learn, but it takes time. So its vulnerability might be thought of as a zero-day vulnerability if this were security. The whole fleet is vulnerable until Waymo has a chance to fix it.

Barnden explained that the whole robotaxi proposition is that every vehicle with the same software would respond exactly the same way, presumably more safely than a human. “So when a Waymo, for example, makes a driving error, it means that the entire fleet would make the same error in that same scenario with the same software. Technically the whole fleet should be stood down whenever any vehicle fails to perform.”

That sort of mass recall would be reputationally disastrous. Barnden said, “So much better for the vehicle to just freeze and a ‘remote assistant’ resolve the issue. And there is the human in the loop to act as the shield.”

Bottom line:

Automakers lost credibility when they failed to articulate what L3 vehicles can and can’t do.

The same automakers are angling for a fresh start by pitching L2++ vehicles built on an L4-ready hardware and software platform.

That “L4-ready” narrative, however, is not so different from Tesla’s FSD pitch that is already riding down a slippery slope where the road sign reads “Oops”.

Related story: