PAC-ACT: Post-training Actor-Critic for Action Chunking Transformers in Precision Industrial Contact Manipulation

Summary

PAC-ACT is a reinforcement learning post-training framework designed for pre-trained action chunking Transformer policies, addressing reliability challenges in precision industrial contact manipulation. Vision-language-action models offer broad generalization but suffer from high inference latency and GPU memory costs, whereas visuomotor action chunking policies are better suited for real-time industrial control. However, such policies, typically trained via behavior cloning, experience distribution shift in contact-rich tasks. PAC-ACT reformulates policy optimization at the chunk level, builds an actor-critic architecture over ACT transitions, and introduces a hybrid behavioral prior constraint. On industrial precision contact benchmarks, the method improves task success rate, contact stability, and force safety while maintaining low latency and low GPU memory footprint. The paper (arXiv:2607.09590) was released on 2026-07-10 by Yujie Pang and Zudong Li.