Key Moments
Chris Gerdes (Stanford) on Technology, Policy and Vehicle Safety - MIT Self-Driving Cars
Lex FridmanKey Moments
Chris Gerdes discusses automated vehicles, balancing technology, safety, and policy, emphasizing voluntary guidance for innovation.
Key Insights
Current vehicle safety standards are slow to adapt to rapid technological advancements like AI in autonomous vehicles.
The Federal Automated Vehicle Policy offers voluntary guidance, including a 15-point safety assessment, for AV development and testing.
Operational Design Domain (ODD) and minimal risk fallback conditions are crucial for defining where and how AVs should operate safely.
Validation methods for AVs can include test tracks, real-world driving miles, and simulation, with no single approach mandated.
Ethical considerations in AVs are approached through engineering principles focusing on risk reduction and societal well-being, not just 'trolley car problems'.
Data sharing, particularly for edge cases, is vital for accelerating AV safety improvements, with aviation's ASIAS system as a model.
THE EVOLVING LANDSCAPE OF VEHICLE SAFETY STANDARDS
Traditional vehicle safety standards, established through processes like the National Traffic and Motor Vehicle Safety Act of 1966, are based on minimum performance requirements with objective tests. However, these rulemaking processes are very time-consuming, often taking seven years or more. This slowness poses a significant challenge for rapidly evolving technologies like deep learning and AI in autonomous vehicles (AVs), where solutions developed today could be outdated by the time regulations are finalized. The current system relies on manufacturer self-certification, which may not be agile enough for the pace of innovation in AVs, prompting a need for new approaches.
THE FEDERAL AUTOMATED VEHICLE POLICY FRAMEWORK
Recognizing the limitations of traditional standards, the U.S. Department of Transportation introduced the Federal Automated Vehicle Policy, offering voluntary guidance rather than strict regulations. This policy encourages manufacturers to voluntarily follow specific guidance and submit safety assessments. It's designed to foster innovation by allowing companies to define their own safety approaches, with the expectation that best practices will emerge over time. This proactive, guidance-based framework aims to balance public safety with the need for accelerated testing and development of AVs, which is crucial for gathering real-world data.
OPERATIONAL DESIGN DOMAIN AND MINIMAL RISK CONDITIONS
A key component of the federal guidance is the concept of the Operational Design Domain (ODD), which requires manufacturers to clearly define the specific conditions under which their AV systems are intended to operate. This includes factors like geographical area, road types, weather conditions, and time of day. Alongside the ODD, developers must define minimal risk or fallback conditions, outlining what the system will do if it encounters a situation outside its ODD or if a system failure occurs. This approach allows for diverse AV designs, from low-speed shuttles to highway-capable vehicles, ensuring they operate within defined safe parameters.
VALIDATION METHODS AND ETHICAL CONSIDERATIONS
The guidance acknowledges various methods for validating AV safety, including test tracks, real-world driving with extensive mileage, and simulation. Each method has limitations; test tracks lack real-world unpredictability, real-world testing may not cover rare edge cases, and simulations must accurately reflect real-world complexities. Crucially, the policy addresses ethical considerations, moving beyond abstract 'trolley car problems' to practical engineering challenges. Manufacturers are prompted to consider how their AVs interact with pedestrians and other road users, using risk reduction principles similar to how automatic emergency braking systems already differentiate between obstacles like vehicles and humans.
THE ROLE OF LEARNING VS. PROGRAMMED RULES
A central debate lies in whether AVs should be programmed with fixed rules or learn from data, mirroring human driving behavior. While human error causes a vast majority of accidents, simply replicating human driving might not achieve the full safety potential of automation. Conversely, purely rule-based systems struggle with the infinite variety of real-world scenarios. The challenge lies in finding a balance, leveraging learning algorithms to adapt to unforeseen situations while ensuring robust safety, and potentially exceeding human capabilities in areas like precision and reaction time, as demonstrated by advanced research vehicles.
ACCELERATING SAFETY THROUGH DATA SHARING AND COLLABORATION
The transcript highlights the immense potential of data sharing, especially for rare, critical 'edge case' scenarios, to accelerate AV safety. By creating shared databases of such events, developers can train AI models more effectively. Inspired by aviation's ASIAS system, where airlines anonymously share safety data, a similar collaborative model could benefit the AV industry. Despite intellectual property and privacy concerns, creating frameworks for sharing anonymized or aggregated data could foster trust, inform regulators, and lead to safer vehicles for everyone, demonstrating a path towards a proactive safety culture and global harmonization of best practices.
Mentioned in This Episode
●Products
●Software & Apps
●Companies
●Organizations
●Books
●Studies Cited
●Concepts
●People Referenced
Common Questions
The US relies on a system of federal motor vehicle safety standards, which are minimum performance requirements. Manufacturers self-certify that their vehicles meet these standards before they are sold, unlike pre-market certification systems in other parts of the world.
Topics
Mentioned in this video
Host of Mythbusters, who took a ride in the autonomous DeLorean, Marty.
Professor at Stanford University studying autonomous car technology and policy, former Chief Innovation Officer at the US Department of Transportation.
IndyCar driver who has helped test and benchmark the performance of the self-driving racecar Shelly.
Actor who appeared in an Audi commercial featuring the self-driving car Shelly.
Actor who appeared in an Audi commercial featuring the self-driving car Shelly.
Author of 'Unsafe at Any Speed,' which influenced the creation of vehicle safety standards in the US.
University where Chris Gerdes is a professor and conducts research on autonomous vehicles.
Department where Chris Gerdes served as Chief Innovation Officer and helped develop federal automated vehicle policy.
US agency responsible for vehicle safety standards and testing.
Part of the US Department of Transportation, responsible for a report on potential barriers to automated vehicles.
A federally funded R&D center that manages the Esaias aviation safety data sharing system.
A committee where Chris Gerdes represented the DOT and advocated for broad availability of AI datasets.
Institution involved in outlining next steps for data sharing and working on the policy in a pilot mode.
A self-driving Audi TT used as an automated racecar by Stanford.
Safety feature advocated for by Ralph Nader in 1965, now standard in modern vehicles.
Safety feature advocated for by Ralph Nader in 1965, now standard in modern vehicles.
A modified DeLorean used by Stanford to demonstrate advanced autonomous driving capabilities beyond human limits.
A car model prominently featured in Ralph Nader's book 'Unsafe at Any Speed' as an example of a potentially unsafe design.
Vehicles used in an automated highway project Chris Gerdes worked on as a PhD student.
A safety system in vehicles that uses different algorithms based on whether the obstacle is a vehicle or a human, highlighting ethical considerations.
A truck platooning firm co-founded by Chris Gerdes, focusing on vehicle-to-vehicle communication.
Company mentioned for its data-driven approach to developing automated vehicle technologies, utilizing data streams from its vehicles.
Company where Chris Gerdes worked on lidar for heavy trucks and suspensions.
Manufacturer that made statements about assuming liability for accidents involving their autonomous vehicles.
A startup company that provided a new electric drivetrain for the modified DeLorean, Marty.
Company mentioned for holding significant data related to self-driving car programs, posing a challenge for data sharing initiatives.
Automaker whose commercial featured the self-driving car Shelly.
Google's self-driving car project, mentioned in the context of the vast amount of data it collects.
Location where the self-driving car Shelly has been driven.
Location where the self-driving car Shelly has been driven.
Racetrack where the self-driving racecar 'Shelly' operates at speeds up to 120 mph.
University where Chris Gerdes worked on automated highway projects as a PhD student.
State mentioned regarding traffic laws like double yellow lines and how they might conflict with automated driving behavior.
Existing US regulations that are minimum performance requirements for vehicles, often requiring objective tests.
A policy developed during Chris Gerdes' tenure at USDOT, providing voluntary guidance for automated vehicle manufacturers.
NHTSA's interpretation of existing rules allowing references to 'driver' to include AI systems in automated vehicles.
Act passed in 1966 that established NHTSA and the federal motor vehicle safety standards.
A machine learning technique that presents challenges for regulators and policymakers in ensuring vehicle safety.
A type of machine learning algorithm that presents challenges for regulators, particularly when behavior is learned.
A classic ethical thought experiment often discussed in the context of autonomous vehicle decision-making in unavoidable accident scenarios.
More from Lex Fridman
View all 546 summaries
311 minJeff Kaplan: World of Warcraft, Overwatch, Blizzard, and Future of Gaming | Lex Fridman Podcast #493
154 minRick Beato: Greatest Guitarists of All Time, History & Future of Music | Lex Fridman Podcast #492
23 minKhabib vs Lex: Training with Khabib | FULL EXCLUSIVE FOOTAGE
196 minOpenClaw: The Viral AI Agent that Broke the Internet - Peter Steinberger | Lex Fridman Podcast #491
Found this useful? Build your knowledge library
Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.
Try Summify free