RL for autonomous driving

Self-driving asks a vehicle to perceive, predict, and act safely in traffic. RL is appealing for the decision-making layer in particular — lane changes, merging, negotiating an intersection — where the right behaviour depends on what the other drivers are about to do.

The state combines sensor-derived features such as nearby vehicles, lanes, and signals; the actions are steering and throttle, or higher-level maneuvers; the reward balances progress, comfort, and safety. Because crashes are unacceptable, RL is trained almost entirely in simulation, often against traffic that itself learns (a multi-agent setting), and then transferred with great care. Many production stacks use RL for sub-problems only, keeping rule-based safety layers wrapped around it.

The long tail of rare events — a child darting out, a ladder fallen on the road — dominates real-world safety yet is nearly absent from data, and one catastrophic failure is one too many. So fully end-to-end RL driving remains largely a research aspiration rather than a shipped product.