Chris Urmson — CEO and co-founder of Aurora Innovation, former CTO of Google's self-driving car team, and technical director for Carnegie Mellon's autonomous vehicle entries in the DARPA Grand and Urban Challenges.
The gist
Chris Urmson traces the technical evolution of autonomous driving from the DARPA Grand Challenge through the Urban Challenge to today's real-world urban deployment efforts. He argues lidar, cameras, and radar are all essential for robust perception, pushing back on the framing of lidar as merely a crutch. He warns that level-2 driver-assistance systems lead humans to over-trust the technology and that their economics diverge from the path to true self-driving. Urmson explains how Aurora aims to demonstrate safety through thorough functional-safety processes, simulation, and on-road data rather than a single metric like disengagements. He predicts large-scale driverless deployment within ten years, starting in moderate-speed urban and suburban environments.
Big reveals
- Urmson responds to Elon Musk's claim that lidar is a crutch, arguing all sensor types (lidar, camera, radar) are essential for robustness.
- He argues level-2 systems are dangerous because humans over-trust them, and their economics diverge from the path to true autonomy.
- He cites people sleeping in their Teslas and Trevor Noah calling a Tesla a self-driving car as evidence of dangerous public misconception.
- Urmson states he told the DMV on the record that disengagements are not a great safety metric.
- He predicts large-scale driverless deployment (10,000+ vehicles) on public roads within ten years.
- He explains Aurora chose urban/suburban driving over trucking because mistakes at low urban speeds are far less catastrophic and enable faster learning.
- When asked what breakthrough would most accelerate autonomy, he says a perfect perception/forecasting model of the next five seconds around the vehicle.
Things worth remembering
- Urmson was technical director of CMU's DARPA team under Red Whittaker and won the DARPA Urban Challenge.
- HD mapping with decimeter-resolution models was the key technology that unlocked the Grand Challenge.
- Multi-beam lidar generating high-resolution mid-to-long-range 3D models was the game-changing innovation for the Urban Challenge.
- Getting centimeter-resolution localization requires care about coordinate systems like NAD83 versus WGS84, accounting for tectonic shifts.
- Urmson cites roughly 37,000 Americans killed on roads in a year as the urgency driving self-driving technology.
- A driver using a level-2 system daily for a month covers only about 1,800 miles, a tiny sample versus the ~85 million miles between fatalities.
- A collision-mitigation system that worked only 50% of the time would still be as transformative for safety as seatbelts.
- Aurora draws on aviation's event pyramid model, statistically linking low-severity events to rare high-severity crashes.