Self-Driving Lessons from Comma AI
summary
An open source level‑2 driving agent openpilot runs today on a <$1,000 Comma Neo device and claims ~10 miles between disengagements now with software updates expected to push to 1,000 miles between disengagements. The stack uses camera‑only perception (seg nets, mono depth, SLAM), dilated causal convolutions and GANs for planning/world models, almost a million miles of video for training, over 50 devices deployed, and concrete safety choices: six‑minute attention timeout and gas/brake disengages.
AI-written summary · about
chapters (11)
- 0:00Introduction to Comma.ai
- 1:55Understanding driving levels
- 4:25Performance metrics
- 6:59Robotics framework
- 8:36Localization and perception
- 12:22Addressing SLAM
- 13:50The planning challenge
- 19:49The power of big data
- 22:36Reinforcement learning
- 25:00Call to action
- 25:44Q&A and closing thoughts
related talks
40mTeaching a Car to Drive Itself by Imitation and Imagination
31mDeep learning for self-driving vehicle technology
35mDingoCar: Perception module Small scale, self-driving car
27mDriving Autonomous Vehicles Forward w/Real World Applications
08mHuman–AI Collab for Reliable Automotive Control Systems
41mAI applications in autonomous driving
52mPhysical AI: Crafting Resilient Systems with Emotional Intelligence
17mWaymo's EMMA: Teaching Cars to Think