Cruise's VP of robotics on SF robotaxi operations

Engineering & simulationreinforcement learning

Cruise uses reinforcement learning to train the autonomous vehicle to behave more cautiously in scenarios where pedestrians might turn around unexpectedly, such as when a hat falls off, improving safety in edge cases.

“We use reinforcement learning to create caution around scenarios like that where you may not know every aspect of why somebody, why that scenario arises … But you do know that in many cases, the pedestrian might turn around. So using reinforcement learning, the AV then learns to behave more cautiously in some of those scenarios.”

Engineering & simulationcontinuous learning machine

Cruise feeds data from rare edge cases, such as people in costumes or debris, through a continuous learning machine to tag and improve the autonomous vehicle's perception and decision-making capabilities.

“We feed them through our data tagging process and then do the continuous learning machine application to improve the AV.”

Engineering & simulationWorld Gen

Cruise built a tool called World Gen to quickly set up simulations for new towns or parts of town, allowing them to test specific scenarios in a simulated environment before real-world deployment.

“we've built the tools to uh, that called, that tool being called world gen, to uh, set that up. And now you can test, uh, all the different scenarios, uh, specific to that uh, world as well.”

Engineering & simulationnatural language processing algorithm

Rashed Haq wrote an NLP algorithm to generate poetic captions for his photography portraits, integrating machine learning into his artistic workflow.

“I used, uh, I wrote a, uh, natural language processing, uh, algorithm, a machine learning algorithm, to give it uh, captions or titles and the titles were very, uh, meant to be very poetic.”