New RL method cuts autonomous driving failures by 5-7% using uncertainty-aware expert advice

A team from Germany (Abouelazm et al.) has developed a novel reinforcement learning (RL) framework for autonomous driving that addresses the fundamental safety-exploration dilemma. Their approach, detailed in a paper accepted at ITSC 2026, uses uncertainty quantification to dynamically trigger expert guidance. When the agent's epistemic (model) or aleatoric (sensory) uncertainty exceeds adaptive thresholds computed from rolling buffers, a human-like expert provides corrective maneuvers. A commitment-cooldown strategy with a stochastic early-stop heuristic ensures the agent experiences coherent driving sequences without becoming dependent on the expert, preserving the exploration budget.

The method integrates expert and agent experiences into a shared replay buffer within an off-policy Implicit Quantile Network (IQN) backbone. Experiments in the CARLA simulator for unsignalized intersection navigation showed that the uncertainty-aware framework improved task success by 5-7% over the standard IQN baseline while significantly reducing collision and off-road failure rates. This demonstrates that risk-sensitive uncertainty coupling with regulated expert advice enables safer, more efficient reinforcement learning for sensor-based autonomous driving policies.