Can Your (ADAS) Car Stop Before Hitting the Kid?

In the era of robotaxis, I assumed that braking when warranted — without human intervention — is a basic safety task long since licked by the automotive industry. But apparently, I was wrong.

(Image source: YouTube video by Mark Rober)

By now, a video by YouTuber Mark Rober, former NASA and Apple engineer, has circulated throughout the automotive industry. I am sure you’ve seen it, too.

This video, using a real-world demo, shows a vehicle equipped with both cameras and lidars automatically hitting the brakes before hitting a child on the road. In several tests, however, a car only equipped with cameras but no lidar — Rober used a Tesla — hits the kid.

Many Tesla fans deemed Rober’s film an unfair assault on Tesla. In response, Phil Koopman, professor at Carnegie Mellon Univ. pointed out on LinkedIn, “This video is about Automatic Emergency Braking, showing that lidar has a natural sensing advantage over camera for AEB approach, especially for pedestrians.”

Koopman added, “This does not mean that the camera can't be made to work (perhaps), but rather that it is a steeper hill to climb to make it work compared to an AEB lidar. And lidar indeed does carry a cost.”

I bring this up because of my current obsession with the National Highway Traffic Safety Administration’s (NHTSA) upcoming standard focused on AEB, FMVSS No. 127.

The rule obligates carmakers to feature AEB and Pedestrian AEB (P-AEB) effective in both daylight and nighttime on all passenger cars and light trucks by September 2029. It also requires the system to hit the brakes automatically up to 90 mph when a collision with a lead vehicle is imminent, and up to 45 mph when the car detects a pedestrian.

My AEB fixation, however, does not stem from the safety agency’s mandate. I was surprised, rather, by the automotive industry’s vehement opposition.

Advancements in sensor technologies are already enabling new models to perceive the road environment much better than ever. Even without Rober’s YouTube, the industry has come to understand that using multimodal sensors — cameras combined with lidars, for example — has enhanced the detection of vulnerable road users and unknown objects.

As a reporter, I’ve covered the full range of sensor technologies over the years. I thought AEB was a solved problem, until Tesla evidently decided to stick to its camera-only guns—as seen on YouTube.

Why isn’t AEB a fait accompli?

In the era of highly automated vehicles, braking when warranted — without human intervention — is a basic safety task long since licked by the automotive industry. But the more I asked about AEB, the more I heard about “phantom braking.”

The AEB issue, as it turns out, doesn’t just hinge on a simple choice of “to lidar or not to lidar.” (Phil Koopman, by the way, gets to the bottom of this topic in his Substack: To Lidar or Not To Lidar )

The phantom braking phenomenon occurs when an AEB system mistakenly thinks it’s going to hit something and brakes unexpectedly. This sudden and unexpected deceleration can cause accidents, especially at high speeds or in heavy traffic.

NHTSA, for example, has been investigating Tesla’s AEB since 2022. The regulators started to study Autopilot’s AEB—involving more than 400,000 vehicles—after drivers reported phantom braking.

Tesla isn’t alone. In January 2025, NHTSA opened a probe into about 295,125 Honda Motor Co. vehicles over reports of AEB systems triggering rapid deceleration inadvertently, increasing collision risks.

In late January, NHTSA also upgraded its probe into Ford’s BlueCruise to an “Engineering Analysis” — a preliminary step before demanding a safety recall.

NHTSA began investigating the BlueCruise safety assist system after two Mustang Mach-E crashes that killed three people. Both times, gong faster than 70 mph (112 kph), the Fords collided at night with cars stalled on travel lanes of controlled-access highways.

During the investigation, NHTSA discovered that a BlueCruise going faster than 62 mph is programmed to ignore stationary objects. Reportedly, Ford made this questionable decision to prevent phantom braking.

Two schools of thought

The obvious questions are how serious the problem of phantom braking is, and how it can be solved. If fully autonomous vehicles are safer than hand-driven cars—as claimed by robotaxi proponents—it should logically follow that the auto industry already knows how to stop a car when it sees that kid in the road.

Asked about phantom braking, Koopman gave me a short summary. “Presumably, Waymo has solved it well enough.” Setting aside the impossibility of “being perfect,” which also applies to human drivers, “it can be solved by more capable sensors and algorithms.”

“For practical purposes,” Koopman sees phantom braking as “a cost issue, not a technology issue.”

Car companies can throw more money at sensors.

In fact, automakers’ demand for sensors is growing. Yole Group predicts the ADAS segment, driven by support for autonomous and safer driving, will likely generate almost $20 billion in revenue by 2029.

But hardware alone is not enough. What’s not clear is whether the industry is ready to invest also in software that solves the phantom braking conundrum.

Michael Brooks, executive director at the Center for Auto Safety, said that automakers know what they must do, but they have yet to commit sufficient resources.

In Koopman’s view, “The industry historically screams about cost until they are dragged kicking and screaming to provide the capability. And then the cost comes down.”

He said: “Providing high-speed AEB and robust pedestrian AEB is clearly feasible with current technology. The challenge … is bringing the cost down to something the industry finds palatable.”

In contrast, Professor Missy Cummings, who focuses on robotics and artificial intelligence at George Mason University, told me, “I do not believe phantom braking is solved, far from it.”

She added, “I also think NHTSA did not do enough research in this area before the rule making was announced.”

In a recent technical paper, she wrote:

“Computer vision false positives, aka hallucinations or phantom braking, need significantly more research, as well as if and how simulations can be used to help identify those conditions likely to result in both missed detections and false positives.”

Cummings also noted, “I completely support the [NHTSA’s] motivation behind mandating P-AEB but right now, it cannot be safely and reliably done with vision alone.”

False positives vs. false negatives

Bryan Reimer, research scientist at MIT’s Center for Transportation and Logistics & AgeLab, also worries about false positives.

“As you try to brake at faster speeds with sensing technology that has not evolved to the degree we would like, you create a higher probability of false positives,” Reimer noted, which are especially perilous at high speeds with following traffic. He said that NHTSA’s “relatively short deadline” for carmakers to solve false positive problems makes compliance “infeasible for most OEMs and suppliers.”

The Center for Auto Safety’s Brooks fully supports NHTSA’s FMVSS No. 127, but he also echoes Reimer.

Brooks said automakers might have to load lots of sensors onto vehicles, creating redundancies and developing a type of verification or sensor fusion that allows a computer to make the right braking decisions. Beyond sensors, software becomes critical for object and event detection, classification and response.

Brooks warned that a phantom-braking solution will be neither cheap nor simple.

Cornered themselves in…
There is one more problem. How do automakers break the news to drivers that AEB and P-AEB installed in ADAS vehicles might not work as well as advertised?

As Phil Amsrud, associate director, automotive at S&P Global Mobility, noted, “Because automakers so enthusiastically embraced self-driving, now any attempt to back away from it” paints them into a corner.

After all, we are living dangerously in an alternative reality today, as Amsrud said. He cited the fact that “Elon Musk is claiming that [a Tesla] car at Level 2+ is capable of Level 4, and we are supposed to take his word for it.”

Balancing act

For OEMs, the phantom braking solution is a happy medium that combines usability, customer experience and safety, according to Benjamin May, CEO of AMX13, a consultancy for mobility and technology.

Explaining that phantom braking often triggers not just surprise among consumers but a flood of complaints, May said OEMs are forced to seek “a compromise that minimizes phantom braking,” sometimes at the expense of safety.

Historically, carmakers dealt with phantom braking by filtering out infrequent cases or adjusting their filters just enough to get their vehicles past New Car Assessment Program (NCAP) tests.

But May said the industry needs “a phased approach,” instead of just cramming for the NCAP test.

May doesn’t think sensors are the limiting factor in phantom braking. The software chain that comes after sensing, he said, requires continuous updates.

Bottom line:

Is phantom braking a solved problem? Hardly.

As long as the auto industry dithers with its vehicles’ hallucinations and false positives (which they tend to adjust more for consumer convenience than safety), this is a phenomenon that’s going to fly under the radar. Carmakers only face up to it when compelled to argue against emerging regulations, like FMVSS No. 127.