The Race Car and the Rulebook: Chris Gerdes on Automated Vehicle Safety

Gerdes begins by noting that he has just returned from a year at the U.S. Department of Transportation and is now speaking in his personal capacity. He wants to connect the deep-learning and neural-network techniques taught in the course to the problems regulators face when they try to ensure vehicle safety. The talk is therefore both an engineering story and a policy story about how automated vehicles are tested, certified, and governed.

He has worked on automated vehicles since 1992, starting with Lincoln Town Cars in an automated highway project as a Berkeley PhD student. He later worked on heavy trucks and suspensions at Daimler-Benz before returning to Stanford, where his group built the steer-by-wire and drive-by-wire electric vehicle P1 and instrumented vintage race cars. He also co-founded Peloton Technology, a truck-platooning company that used vehicle-to-vehicle communication to shorten following distances.

The best-known Stanford vehicle is Shelley, an automated Audi TT that can reach about 120 mph on Thunderhill Raceway, limited mainly by the length of the straight. Watching from outside is occasionally disconcerting because nobody is in the car, although Gerdes says it looks calm from inside. Shelley now exceeds the performance of everyone on the development team and can beat the track president; an IndyCar driver is faster, but by only about a second on a one-minute-and-25-second lap.

Gerdes explains that Shelley was built around a physics model: the car calculates how to brake into the next corner and how much grip it believes it has, using force equals mass times acceleration. The core approach is not a learning algorithm, although the team has added learning on top. The gap between the car's performance and the best human performance is where learning matters.

The best human drivers instinctively push the car to the limit rather than prejudging what the limit is. Learning algorithms help Shelley estimate how much friction is available in a particular corner and how that changes as tires and track warm from morning to afternoon. Gerdes argues that the same capability matters for safety: on an unknown road, the car should be able to use all available tire-road friction to make the safest possible maneuver.

Racing offers a useful discipline because drivers know that to finish first they must first finish, so speed must be paired with accident avoidance. The Stanford goal is not to make every drive feel like a qualifying lap but to let the car avoid any accident it is physically capable of avoiding. That requires understanding and exploiting limits that a cautious, preprogrammed system might never approach.

In the U.S., the Federal Motor Vehicle Safety Standards are minimum performance requirements, each with an objective test. No federal agency tests vehicles before they are sold; manufacturers self-certify that they meet the standards, and NHTSA can later purchase and test vehicles. Most of the rest of the world uses premarket certification, and aviation requires certification before sale, but cars do not.

Gerdes traces the system to Ralph Nader's 1965 book Unsafe at Any Speed, which argued that the auto industry had suffered a failure of imagination rather than a failure of engineering. Nader advocated airbags and anti-lock brakes long before they arrived, at a time when many people assumed that crashes above about 40 to 45 mph were inherently unsurvivable. Public outcry led to the National Traffic and Motor Vehicle Safety Act of 1966 and the standards system that followed.

Rulemaking is painfully slow: Gerdes says the minimum optimistic time for a new standard is about two years, and realistically it is closer to seven. That timeline is incompatible with automated vehicles, because rules started seven years ago would govern systems that have already changed. He also argues that the original premise of a sluggish industry has flipped: today there is no failure of imagination among startups, automakers, and technology companies.

A Volpe report found little at the federal level that prevents an automated vehicle from going on the road if it meets the existing standards, since no standard specifically addresses automation. NHTSA also interpreted references to a driver as potentially referring to the automated system. The exceptions are designs without a steering wheel or pedals, which need an exemption because those requirements reflect older assumptions about human-driven cars.

The policy problem is to encourage testing so developers can gather real-world experience while preserving the government's responsibility for safety. NHTSA can recall systems that create an unreasonable risk, but Gerdes wanted a more proactive approach. The September federal automated vehicle policy created voluntary guidance and asked manufacturers to submit a safety assessment letter explaining how they addressed specific areas. The aim was not to prescribe methods but to start a conversation from which best practices could emerge.

The most relevant part for vehicle design is the 15-point safety assessment. It begins with operational design domain and minimal-risk or fallback conditions: developers must define where the system is supposed to work, how they ensure it stays in those conditions, and what happens if it does not. A human-driven car might ask the driver to retake control, while a vehicle with no steering wheel needs a different fallback. The guidance also asks how the consumer has been educated about what the system can and cannot do.

Validation methods are left open, so developers can combine test-track maneuvers, real-world miles, simulation, and analysis. Gerdes notes that real-world testing in one place may not transfer cleanly to another, while simulation can replay and perturb scenarios far more extensively than road testing. The guidance is neutral about the mix, and companies with different data access will choose different paths: Tesla can run ideas silently on customer vehicles, while others may rely more on simulation or track testing. It also anticipates over-the-air software updates.

Gerdes says ethical considerations often evoke the trolley problem: a car can plow into ten people or divert and kill its driver. He is skeptical of framing automated-vehicle ethics as an impossible algorithmic morality problem. Engineers should work with philosophers, he argues, but not adopt the trolley problem too literally. The practical questions are different: who is responsible for the brakes, why is there no backup system, and why is the car heading into a group of people without the ability to stop?

Reframed as engineering, the trolley scenario produces design actions rather than a moral ranking. The engineer adds emergency braking, then a backup for the emergency braking, then perhaps diverts the trolley onto a side track. Knowing that the side track might be used, the engineer puts up a fence and a warning sign. If someone ignores the sign and is hit, the engineer may feel differently, but the solution came from reducing risk rather than deciding who deserves to live or die.

Gerdes also points to broader ethical choices that manufacturers already make. How close should a car get to pedestrians and bicycles, and how much care should it show to other road users? Automatic emergency braking systems can use different algorithms depending on whether the obstacle is a vehicle or a human, braking harder for the human because the potential harm is greater. The guidance's ethical section is meant to open discussion of these choices and move toward best practices that make all road users feel cared for.

A second ethical tension is between safety, mobility, and legality. A van may be parked illegally on a double-yellow line, and California law does not clearly permit crossing the line to get around it, yet most human drivers would cross. Drivers also routinely cross double-yellow lines to give cyclists room, which is not what the vehicle code says. If automated cars follow the letter of the law, they may create rolling roadblocks and drive in ways that no human wants to share the road with.

The design question is whether automated vehicles should drive like humans or like robots executing the law. Fixed rules are one option, learned behavior from human drivers is another, and a balance may be possible by weighting objectives analytically. Gerdes notes that all approaches have limits in the extreme: Shelley showed that without learning from experience, a car does not use all the data available to it. No one can preprogram every scenario, so interpolation and learning are unavoidable.

The counterargument is that human error is a primary factor in 94 percent of accidents, through failures of judgment or perception. If automated vehicles simply imitate humans, they learn only how well humans do and leave much of the car's potential unused. Gerdes frames this as an ongoing discussion for both developers and policymakers: what should be learned, what should be programmed, and how to avoid leaving capability on the table. The answer will shape both performance and safety.

He illustrates the potential with Marty, a DeLorean that his lab converted to an electric drivetrain after fixing its weak brakes, underpowered engine, and lack of power steering. When programmed to drive itself, Marty can execute a perfect doughnut around a cone, launch sideways through a gate, and orbit points with a precision the team believes no human can match. The car follows a trajectory akin to orbital mechanics, and even as its tires disintegrate, it can repeat the path with tire tracks nearly on top of the previous run. Beyond the friction-saturation limit assumed in ordinary driving lies a class of emergency trajectories that could be valuable in some situations.

Gerdes sees large potential in sharing data about edge cases. Training a neural network to handle extreme situations is easier when the training set contains those situations, and a shared database of weird road events would give every developer a head start. The challenge is to balance intellectual-property concerns and individual privacy with the safety benefits. He argues that academia can help broker the discussion by moving beyond abstractions and asking what information would actually help regulators and companies.

Aviation already has a model: the ASIAS system began with four airlines sharing safety information through MITRE, a federally funded R&D center, and now includes about forty airlines. Participation is valuable enough that companies work hard to get back in if they are removed. The data is anonymized, companies receive an assessment of their safety record and can compare themselves to the group in the abstract, but no identifiable airline is ranked and the system is not used for enforcement. Over time trust grew and airlines shared more data.

Gerdes connects this to a broader AI point: the field has advanced when good datasets became broadly available, and transportation needs the same. As the DOT representative on a White House subcommittee on machine learning and AI, he pushed recommendations that such datasets be made widely available. The automated-vehicle policy's vision was a proactive safety culture, not premature standards, with government involved early in discussions and the U.S. continuing to play a leading role in testing and development.

The framework is meant to evolve: guidance generates conversations, conversations generate best practices, and best practices can eventually become standards or law. NHTSA will not develop all of those practices alone; companies and universities will create many of the solutions. Gerdes ends the talk with a challenge to students to think not only about the technology but about how to communicate it to regulators and where the gaps are that research, startups, or government work could fill.

Asked whether automated vehicles could allow passive safety equipment to be rolled back, Gerdes says yes, eventually, but with caveats. Passenger cars use roughly 90 percent of their energy to move the vehicle rather than the people inside, partly because crashworthiness standards add mass. If vehicles are designed not to crash, or operate only in low-speed urban environments separated from pedestrians, much of that mass could be removed. That could make transportation options comparable to cycling from an environmental standpoint, although it would require limiting the operating environment or relying on far-future techniques.

On liability, Gerdes says the question is partly open but not entirely new: courts have long sorted out responsibility for new technologies, and U.S. joint-and-several liability can assign different portions of responsibility to different players. Volvo and Google have said they would expect to be liable if their automated vehicles were involved in accidents. The helpful change would be more harmonization across states, since insurance and fault rules are set state by state. On public acceptance, he adds that people may tolerate human error more readily than machine error, especially when a failure looks visually striking, so automated systems may be held to a higher standard than the overall fatality statistics alone would suggest.

At the G7 transportation ministers meeting, Gerdes discussed how countries with different certification systems might still share best practices on testing. Since manufacturers operate globally, a common set of standard tests could be translated into different regulatory environments. He also returns to data sharing: the practical path is to start small with high-value information that helps public policy without exposing proprietary advantage, such as interference data from vehicle-to-vehicle communications. DOT's next steps included pilot work on sharing map data, edge-case accident data, and simulation capabilities.

Asked whether urban and rural environments require standard datasets, Gerdes returns to operational design domain: a company that only intends to operate in a campus or a small urban area should explain how it handles eventualities there, not everywhere. Common datasets become more interesting as design domains expand nationwide, but no finite dataset captures every eventuality, and companies can design to a test. He also says an open-source car is not obviously barred by the guidance, which applies to universities and other groups too, but someone must sign the safety assessment and take responsibility for the version of software actually used. The 15 points are not meant to be overly restrictive; any group putting real people at risk should have thought through them.