These are the slides of my two-part lecture on reinforcement learning for autonomous driving at the ELLIS Summer School on Autonomous Driving, Barcelona. Both decks are interactive Reveal.js presentations:
- Arrow keys or hjkl to move between slides
- F for full screen
- Esc or O for an overview.
Part 1: RL background
Quick overview over the RL foundations:
- Why do we need RL: covariate shift and its experimental evidence.
- From first principles to PPO and GRPO, including brief intros to value functions and GAE.
- An overview over the broader RL landscape: exploration, offline RL, model-based RL and POMDPs.
- What makes RL powerful and fragile: the data moves with the policy.
Part 2: The research frontier of RL for Autonomous Driving
Mostly structured as Problem -> Existing Solutions -> Examples, including topics such as:
- Reward design
- Sim2Real gaps
- Traffic modelling
- Challenges of RL for E2E driving
- Imitation-to-RL gaps
Open Part 2: The research frontier
The decks are rendered with Quarto.