Safe Decision-Making via Adaptive Causal Representation for Autonomous Driving.

Area of Science:

• Autonomous Driving Systems
• Machine Learning
• Reinforcement Learning

Background:

• Offline reinforcement learning (RL) for autonomous driving struggles with out-of-distribution (OOD) scenarios due to deployment shifts.
• Unseen road geometries and varied driving behaviors can lead to unreliable decisions from offline-trained policies.

Purpose of the Study:

• To develop a safe offline-to-online decision-making framework for autonomous driving.
• To enhance policy reliability and generalization in evolving, OOD deployment conditions.

Main Methods:

• An adaptive causal transformer (AC-Transformer) learns a causal representation from offline driving trajectories, modeling state-action-reward-cost influences.
• Bisimulation regularization organizes the representation into a consequence-consistent latent structure.
• A phase-adaptive objective (PAC) progressively refines the representation during online deployment.

Main Results:

• The proposed method demonstrates stronger generalization across diverse road structures and driving styles.
• It effectively balances safety and efficiency in dynamic driving environments.
• Outperforms eight representative baselines in five distinct driving scenarios.

Conclusions:

• The adaptive causal representation framework enhances the safety and reliability of autonomous driving decisions.
• The method provides robust performance under deployment shifts and OOD scenarios.
• It offers a promising approach for real-world autonomous driving applications.