UA2C: Uncertainty-Aware Adaptive Action Chunking for Offline-to-Online Decision-Making in Mixed Traffic

Autonomous driving requires sequential decision-making under partial observations and heterogeneous interactions with human-driven vehicles. While action chunking provides a useful temporal abstraction by predicting multi-step future actions, fixed-length chunk execution can be unreliable when later actions are conditioned on stale observations. We propose $\textbf{UA2C}$, an uncertainty-aware adaptive action chunking framework for offline-to-online reinforcement learning (RL) in mixed traffic. UA2C first learns a flow-matching chunk policy from offline driving data and then refines the policy through online interaction. To account for behaviorally diverse surrounding vehicles, UA2C incorporates a driving-style inference module that augments the policy with local behavior context. During execution, UA2C estimates uncertainty from sampled action chunks and executes only a reliable prefix before replanning. Experiments show that UA2C improves offline reward and control smoothness over a one-step baseline, and further improves online performance over fixed chunk execution.