Teaching a Car to Drive Itself by Imitation and Imagination
summary
ChauffeurNet trains a convolutional RNN on a top‑down 80×80 m representation and 26 million examples to output waypoint trajectories; open‑loop metrics look fine but closed‑loop reveals biases. Orientation bias is fixed by randomized heading offsets, past‑motion bias by pose dropout, and tiny initial motions by adding a waypoint‑index input. Recovery comes from randomized position perturbations; collision avoidance uses an overlap loss applied to those synthetic collisions; co‑training predicts other agents and a road mask, and imitation dropout frees exploration.
AI-written summary · about
chapters (9)
- 0:00Start
- 3:29Self Driving Software Stack Mid-to-Mid Driving
- 9:25Chauffeur Net Model
- 31:38Stop Sign Ablation Test
- 32:30Vehicle Ablation Test
- 33:35Passing a Parked Car
- 34:15Recovering from a Trajectory Perturbation
- 34:42Slowing Down for a Slow Car
- 36:59Chauffeur Net: Current Issues
related talks
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
41mSelf-Driving Lessons from Comma AI
31mDeep learning for self-driving vehicle technology
42mAI and Self-Driving Cars
05mInterview
17mWaymo's EMMA: Teaching Cars to Think
30mBuilding, designing, teaching and training simulation environments for Machine Learning
04mInterview