Interview
Chris Urmson: Self-Driving Cars at Aurora, Google, CMU, and DARPA | Lex Fridman Podcast #28
- Aurora projects a large-scale autonomous vehicle deployment within 10 years, potentially scaling to 10,000 vehicles, with timelines potentially accelerating if a compelling user experience emerges or if a perfect model of vehicle behavior for the next five seconds is achieved.
- Level 2 and Level 3 driver assistance technologies are forecast to diverge from full autonomy paths due to distinct economic constraints, specifically the acceptable error rates of "false negatives" in assistance systems versus the fatal nature of such errors in self-driving applications.
- The company anticipates limited mass-market immunity to L2 overtrust risks, noting that even technology-savvy users may drift into overconfidence within one to two years due to the compelling nature of personal experience.
- Aurora plans to prioritize solving full autonomy in moderate-speed urban and suburban environments to enable faster learning with lower risk, expecting highway capabilities to naturally follow as a consequence of resolving urban challenges.
- Level 4 autonomy is expected to be economically viable without relying on the cheapest sensors, with LiDAR costs predicted to decrease substantially over time despite remaining more expensive than CMOS-based imagers due to manufacturing scalability differences.
- The market is projected to prioritize a $500 sensor suite capable of operating an autonomous system over a $50 alternative that fails to function, prioritizing safety and sustainable scaling over minimum cost.
- Safety validation plans include engaging with regulatory bodies like NHTSA to verify rigorous functional safety processes through simulation, unit testing, and on-road data, while developing specific metrics for traffic lights and turns to compare failure rates against human performance.
- Aurora intends to attract top engineering talent by focusing on deep problem-solving and investing in infrastructure and machine learning tools rather than constant demonstration cycles.
- Long-term value propositions are expected to become "mundane" and invisible to users, similar to electricity, enabling safe life experiences without active attention from the driver.
- Remaining challenges, such as game-theoretic interactions with pedestrians for "nudging," are characterized as future human-computer interface design problems rather than fundamental algorithmic unsolvabilities.
- Risks associated with initial user hesitation and lower confidence in vehicle reactions are not expected to create chaos, as human safety instincts generally prevent individuals from exposing themselves to vehicle collisions.