Deep learning for self-driving vehicle technology
summary
Uber ATG describes moving from a modular stack to a single trainable system: a joint detection+prediction network that outputs multimodal future trajectories, plus emerging interpretable neural motion planner architectures that emit intermediate perception and planning representations. Lidar runs at about 10 Hz, HD maps are used, and sensor simulation (lidar and camera) is trained from logs to run hundreds of thousands of scenarios; vehicle-to-vehicle intermediate messages cut simulated perception errors by 70% decrease in perception error. Runtime work focuses on GPUs and lower-precision models.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:43About Inmar
- 1:33Why selfdriving
- 3:40How selfdriving works
- 4:49From sensors to controls
- 6:22The autonomy stack
- 6:57Perception
- 15:27Perception and project prediction
- 16:26Neural motion planners
- 17:28Vehicletovehicle communication
- 19:16Simulation
- 21:57Cameras
- 23:04Putting it on the car
related talks
13mPyTorch at Uber
49mUsing Deep Learning for Perception in Autonomous Systems: A Perspective from Control Theory
08mUnified Perception: Models for Autonomous Driving
05mInterview
27mDriving Autonomous Vehicles Forward w/Real World Applications
42mAI and Self-Driving Cars
04mInterview
19mAI Trends in Intelligent Transport Systems