Predicting the Unpredictable Cruise’s Continuous Learning Machine
summary
Cruise builds a Continuous Learning Machine that deploys a simple model, logs every 10 Hz prediction, finds surprising mistakes (U‑turns, lane cuts, parking‑garage cuts), auto-labels the future from sensor traces and re‑trains on those hard examples in a loop. Key pieces: self-supervised auto-labeling to turn future observations into labels, active error sampling to focus long‑tail maneuvers, and multimodal prediction outputs that represent multiple likely futures. System runs at scale (powering ~2M autonomous miles in SF) with safeguards like mixed‑distribution sampling and conservative update rates to avoid drift.
AI-written summary · about
chapters (6)
related talks
21mML Infrastructure for Autonomous Vehicles
19mThe Data Science behind Cruise’s Robotaxi Fleet
39mChanging Realities - The Final Frontier for Machine Learning?
31mDeep learning for self-driving vehicle technology
40mTeaching a Car to Drive Itself by Imitation and Imagination
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
41mKEYNOTE: Image processing, AI, and autonomous systems