Over the holidays, the media outcry over Waymo’s massive robotaxi traffic jams a week ago—triggered by San Francisco’s citywide blackout—has somehow morphed, thanks to Waymo’s spin-doctored postmortem, into a success story:
Waymo’s self-driving system, designed to treat dark traffic signals as four-way stops, performed that baseline performance across most of the city. Its fleet “successfully traversed more than 7,000 dark signals,” according to Waymo.
The traffic jam stemmed from “confirmation protocols” that Waymo had established “out of an abundance of caution.”
Each robotaxi followed protocol and “requested a confirmation check to ensure it makes the safest choice.” The result was “a concentrated spike in these requests” during the outage. Waymo explained that this created a backlog that, in some cases, led to response delays contributing to “congestion on already overwhelmed streets.”
This, of course, is Waymo’s style of crisis management.
The revised narrative stirred widespread pro-Waymo sentiment among social media influencers. The postmortem gave the press opportunity to pump out fresh stories.
The social media consensus was to: “Give Waymo a break.” Influencers praised Waymo for “learning lessons” and said, “It wasn’t as bad as some of the social media posts made it look like.” They insisted that “the traffic jam wasn’t caused by the failure of Waymo’s self-driving ability, but this was a result of a massive ‘system-wide failure,’” and “poorly maintained PE&G’s infrastructure.”
Absent from all the hindsight, however, is a closer look at what “teleoperation” exactly entails, as it is deployed by robotaxi companies without guidance or regulations by any authorities.
So little is publicly known
Local and federal agencies know very little about teleoperation of robotaxis.
No one seems to know, for example, how many teleoperators are deployed per how many robotaxis or where “call centers” for teleoperators are located. Nor is there information on the criteria for hiring “teleoperators,” or the minimum latency for self-driving remote operations, or what sort of telecommunication network performance is required for handling a spike in calls?
Above all, is it too much for anyone to ask whether there is such a thing as an industry safety standard for the teleoperation of self-driving cars?
Lacking a formal standard, who’s in charge of setting up best practices necessary for safe teleoperation?
Waymo appears to claim that it has “been there, done that.”
In a November press release, Waymo claimed that an audit by an independent assessor “confirmed the adherence of Waymo’s policies and practices with the industry best practice on Remote Assistance Use-Cases produced by the AVSC consortium.”
Great.
If so, what’s in the document?
Waymo’s release helpfully pointed out that TÜV SÜD’s audit was carried out against the AVSC Best Practice, ADS Remote Assistance Use Case AVSC-I-04-2023.
I asked Phil Koopman, safety expert and professor emeritus at Carnegie Mellon University, to help me break down the verbiage.
What’s in the document
Downgrading the AVSC document from “a standard” to “a guideline,” Koopman explained that the Automotive Vehicle Safety Consortium’s membership includes Waymo. Coincidently, the guideline emerged around the same time that Cruise, GM’s self-driving car service, was under scrutiny after one of its robotoxis dragged a pedestrian for 20 feet after a hit-and-run.
The role played by Cruise’ teleoperators was a focus of the investigation.
The industry’s prevailing understanding then was that Cruise robotaxis required frequent human intervention from remote operators. Reports indicated that the driverless cars needed help every 2.5 to 5 miles.
Although Waymo also deploys teleoperators, the company strove to distance itself from Cruise, stating that its remote support provides high-level “advice” rather than “direct control.”
The AVSC teleoperation document appears to reflect the need of the AV industry at that time to make a distinction without a difference between remote assistance and remote driving.
Waymo stressed that “we don’t have remote drivers, we have remote assistance.”
AVSC’s document is only 25 pages. Everyone ought to read it.
The consortium’s teleoperation guideline, heavily focused on taxonomy, spends most of its ink in describing differences in function between remote drivers and remote operators’ operations.
Explicit in the document is that its scope is limited to “Remote Assistance (RA) for Automated driving system (ADSs) only, and it excludes remote driving, dispatching, and customer support to occupants or other road users.”
Koopman noted, “This document defines what it means to have remote assistance and why it’s not remote driving.” Although he sees “nothing wrong with it” as a foundation, Koopman said the problem is that the document is incomplete.
What’s NOT there
The AVSC document says nothing about fleet operations. “This document is about a single interaction with a single car for a single event,” Koopman noted. “It’s a guide to execute a call center transaction at a remote operation center.”
Furthermore, “this is not even a call center standard,” he said, “because it doesn’t talk about workload management at call centers.”
Koopman asked: “Does the document talk about handling loss of link?” Answering his question, he said, “No … This is not a safety standard.” Noting that the document doesn’t identify hazards, Koopman reminded, “If you don’t have a hazard log, it’s not safety engineering.”
The sin Waymo committed in its press release is creating the impression that TÜV SÜD has signed off on the safety of Waymo’s remote operations.
In reality, TÜV SÜD only audited Waymo’s fleet response against an AVSC document that isn’t a safety standard.
Earlier this week, I contacted Bijoy Bhaskaran, director of mobility at TÜV SÜD, asking what he sees as a good practice to implement safe, effective teleoperations. Bhaskaran did not respond by press time.
Missy Cummings, professor at George Mason University, originally wrote on LinkedIn that TÜV SÜD’s “independent” audit of Waymo remote operation can’t be trusted. Noting that “power failures are a known issue, and have happened to Waymo before,” she asked the assessor, “How could you miss something so obvious?”
When I reached out to Cummings this week, she responded, “I assumed the independent audit was based on information not in the public domain. If they did their audit based on the AVSC document, that would be incredibly superficial and basically a joke.”
Remote operation regulations?
Given Cummings’ experience as a senior safety advisor at the National Highway Safety Agency (NHTSA), I asked which aspects of remote operations (or “call centers for AVs”) that NHTSA should examine for potential regulation.
In her opinion, “We need federal regulation that ensure remote operators are in the United States, with US [drivers’] licenses, and that meet the same alcohol and drug testing requirements that airline dispatchers have to meet.”
She added, “Companies should have to have emergency backup remote ops centers as well as contingency plans for power loss that could affect communications. And companies should provide detailed hazard analyses that examine other communication failures as well as issues with latency.”
She would like to see that “any company that is publicly traded should have to report the numbers of remote operators as well as the vulnerabilities that such operations introduce.”
Academic research?
Absent in the teleoperation discussion is academic research. Cummings, for instance, wonders if there is any empirical evidence that suggests the minimum latency for self-driving remote ops.
She noted, “There are notably no standards around latencies which is a major and intentional oversight.”
Cummings referenced a report she wrote for NHTSA while she was the senior safety advisor there. In that report, “Issues in Teleoperation of Autonomous Vehicles,” she not only defines teleoperations, but also lays out “mitigating teleop risks” along with specific recommendations.
In that report, she wrote:
Teleoperation is not new and many different standards exist for the design of various teleoperated devices like those in industries like manufacturing and mining. Such standards help industry mitigate risk but unfortunately, for the most part automotive standards are silent on design and testing guidelines for teleoperated systems6 and many in industry have called for increased guidance and regulation on this point. SAE J3016 recognizes that this is an important topic, but only goes so far as to define remote driving and remote assistance and says nothing about in-vehicle or in-situ teleoperation. SAE J3018 statesthat they may address these issues in a future revision. UL 4600 recognizes the modes of teleoperation/remote driving/remote assistance and identifies them as a significant source of risk that requires mitigation (11.2.5.2), but beyond this, makes no recommendations. Thus, standards are not going to provide any meaningful guidance.
What’s going on in California
To Cummings’ dismay, “The California DMV recently proposed new regulations that notably allow self-driving companies to locate their remote operations centers outside the county and also dropped the requirement to require remote operators to have a license in the US,” she noted. “These were on their website last week.”
Cummings said, “This is in direct opposition to what I advised them. My opinion is that Waymo convinced CA DMV to drop the US requirements. This is an amazingly bad idea, especially in light of the power outage debacle.”
Despite Waymo’s remarkable progress in robotaxi technology, it doesn’t appear to recognize that building social trust takes more than well-crafted press releases.
Related story:
Hey, Waymo! When The Big One Hits, What’s Your Plan?
A domino-effect sequence of traffic jams triggered by Waymo robotaxis during Saturday’s massive blackout in San Francisco has prompted safety experts to wonder whether Waymo is prepared when the long-feared “Big One”—earthquake—hits the city.Thanks for reading Junko's Tech Probe! Subscribe for free to receive new posts and support my work.
Honestly anybody who's familiar with safety audits, esp. in that domain, knew from the beginning that the TÜV Assessment was only covering a small bit. And Waymo played the usual game of saying loud "audited" without providing the footnotes clearly. Marketing as usual.
Still tbh Cummings reaction is really not professional: she didn't do her homework, didn't check what was audited and has no problem "calling out" TÜV without checking the facts and saying what they did was a joke. They did what they were tasked for and Waymo did their advertisement. Cummings just didn't bother to check. It's not the first time she provides such ungrounded statements. I think you should be careful with that.
Considering the necessity of "teleoperators" (a word that could only be coined by a geek in his mother's basement), shouldn't robotaxis be referred to more accurately as grounddrones?