This paper introduces PAC-ACT, a reinforcement-learning post-training framework designed to enhance pretrained Action Chunking Transformer (ACT) policies for industrial precision-contact tasks. The method reformulates policy optimization at the chunk level and utilizes an ACT-transferred actor-critic architecture with a hybrid behavior-prior constraint.
- PAC-ACT preserves the pretrained action distribution during online fine-tuning using a hybrid behavior-prior constraint.
- Experiments on industrial benchmarks show improvements in task success, contact stability, and force safety while maintaining low latency and GPU-memory usage.
- On the Contour task, the framework reduces peak contact force and decreases the proportion of force readings above 60 N by 46 times.
- Sparse-reward ablations demonstrate that the behavior-prior constraint enables effective exploration under randomized initial poses.
The authors consider this significant because it addresses distribution shift in contact-rich tasks, allowing vision-action chunking policies to achieve reliable robot control without the high inference latency and memory costs associated with vision-language-action models.