The long road to driverless with Aurora's Chris Urmson (Live at HumanX) | Equity Podcast

In this talk

1. 0:59

   From technical milestone to scaling

   Urmson traces his 24 years in autonomy from Carnegie Mellon’s DARPA Grand Challenges and mining trucks to Google’s self-driving program and Aurora. He says Aurora has operated driverless trucks in Texas, New Mexico, and Arizona, logged more than 250,000 driverless miles, and is moving from a handful of trucks toward hundreds.

2. 4:21

   Supply, customers, and regulation

   Urmson says technology and safety were the main historical bottlenecks; Aurora is now constrained by hardware supply. He says second-generation hardware with International LT trucks is intended to support about 1,500 trucks, with Volvo hardware planned to enable tens of thousands annually. Aurora operates where permitted, and Urmson says California rules and a federal framework could shape expansion, while the company can serve an estimated 50 billion Sun Belt vehicle miles under current rules.

3. 9:18

   Why trucking, and what customers gain

   Urmson says customers prioritize safety, then value higher truck utilization and fuel savings. He says a truck costing about $150,000 can be used roughly twice as much when it is not limited to a human driver’s 11-hour day, and Aurora expects fuel use to fall by 14 to 34 percent. He says trucking offers a larger market and stronger unit economics than ride-hailing, and credits Aurora’s long-range lidar with making freeway trucking viable.

4. 18:10

   Safety cases and verifiable systems

   Urmson says safety-critical products require time, careful processes, and trust. Aurora’s safety case has five pillars: proficiency, fail-safety, resilience, continuous improvement, and trustworthiness, supported by about 450 pieces of evidence. He also describes Aurora’s verifiable AI approach, which breaks the driving problem into parts so the system can be tested and constrained.

5. 16:37

   Applying autonomy beyond trucks

   Urmson says trucking alone could make Aurora successful, while its physical AI capabilities may later extend to box trucks, robotaxis, mining, agriculture, drones, or humanoid robotics. He says those choices remain open. He also argues simulation and real-world testing must complement each other: field data helps validate and improve simulations, whose limits need to be understood.