Explore a keynote presentation on developing robust driving policies for autonomous vehicles, delivered at the CVPR Workshop on Autonomous Driving. Delve into two cutting-edge approaches that achieve state-of-the-art performance in the CARLA simulator. Discover a novel framework for situational driving policies that adapts to diverse scenarios, resulting in a 98% success rate on the CARLA driving benchmark. Examine the challenge of covariate shift in imitation learning and learn about a new technique that improves generalization by sampling critical states and using a replay buffer. Analyze the performance results on various CARLA benchmarks, including the NoCrash benchmark with dense traffic conditions. Gain insights into the importance of mixture models, task-based refinement, and emergent driving modes in developing robust autonomous driving systems.
Overview
Syllabus
Intro
Conditional Imitation Learning
Sunny, light traffic
Sunny, heavy traffic
Rain, heavy traffic
Sunset, heavy traffic
Inspiration: World Models
Learning Situational Driving
Importance of Mixture Model and Task-based Refinement
Emergent Driving Modes
Results on CARLA Benchmark
Results on CARLA NoCrash Benchmark
Results on AnyWeather Benchmark
CILRS: Collision, infraction
LSD: No collision, proper braking
Formal Definition of Imitation Learning General Imitation Learning
Challenges of Behavior Cloning
Experiment by Held and Hein
Distribution over Driving Actions
Dagger with Critical States and Replay Buffer
Evaluation
Infractions Analysis
Training Variance
Qualitative Results
Summary
Taught by
Andreas Geiger
