Adaptive Action Chunking for Robotic Imitation Learning

Action chunking strategies in robot imitation learning struggle to dynamically balance between long-range motion efficiency and short-range operational precision due to their fixed planning horizon. This paper presents an Adaptive Action Chunking framework that enables robots to dynamically predict the optimal action chunk length based on real-time visual context. We design an end-to-end dual-branch network comprising a shared visual encoder, a parallel action prediction head, and a chunk-size prediction head. Experiments on two real-world bimanual robot manipulation tasks (transport-and-place and flip-and-handover) demonstrate that the method autonomously derives two distinct intelligent strategy patterns—phase-aware switching and sustained high-frequency adjustment—in response to task uncertainty. It significantly outperforms fixed-chunk baselines in both success rate and efficiency. Ablation studies confirm that the performance gain stems from the adaptive decision-making mechanism itself.