Uncertainty-Aware Expert Guidance for Safer Autonomous Driving RL
A novel framework for reinforcement learning in autonomous driving enhances safety during exploration by utilizing expert advice activated by uncertainty thresholds. This approach, outlined in arXiv:2605.30576, incorporates adaptive thresholds from rolling buffers to trigger guidance when either epistemic or aleatoric uncertainty is elevated. To manage the duration and frequency of advice, a commitment-cooldown strategy with stochastic early stopping is implemented, minimizing the risk of over-dependence. Experiences from both experts and agents are integrated into a shared replay buffer, supported by an off-policy implicit quantile network (IQN) backbone. Testing within the CARLA simulator demonstrates that this method surpasses the IQN baseline, effectively decreasing collisions and off-road events while preserving learning efficiency.
Key facts
- Framework uses uncertainty-aware expert advice to guide exploration in RL for autonomous driving.
- Advice triggered when epistemic or aleatoric uncertainty exceeds adaptive thresholds from rolling buffers.
- Commitment-cooldown strategy with stochastic early-stop heuristic regulates advice duration and frequency.
- Expert and agent experiences combined in shared replay buffer within off-policy IQN backbone.
- Experiments in CARLA simulator show method outperforms IQN baseline.
- Aims to reduce collisions and off-road driving during exploration.
- Published on arXiv with ID 2605.30576.
- Addresses inherent unsafety of exploration in autonomous driving RL.
Entities
Institutions
- arXiv