Summary
PAC-Act is a reinforcement learning post-training framework for pretrained action chunking Transformer policies in robotic manipulation, proposed by Yujie Pang and Zudong Li (arXiv:2607.09590). The work targets precision industrial contact manipulation, where policies must remain reliable under pose perturbations and contact force constraints. While vision-language-action models offer broad generalization, they suffer from high inference latency and GPU memory costs, making compact vision-action chunking policies preferable for real-time industrial control. However, policies trained purely with behavior cloning experience distribution shift in contact-rich tasks. PAC-Act addresses this by reformulating policy optimization at the chunk level, building an actor-critic architecture tailored to ACT-style transitions, and introducing a hybrid behavior prior constraint to keep the learned policy close to the pretrained one. On industrial precision contact benchmarks, the framework improves task success rate, contact stability, and force safety while retaining low inference latency and low GPU memory footprint.
PAC-Act: Post-Training Actor-Critic for Action Chunking Transformers
Research area: Robotics / AI
Authors: Yujie Pang, Zudong Li
Published: 2026-07-10
arXiv: 2607.09590
Abstract
Precision industrial contact manipulation requires reliable robot policies under pose perturbations and contact force constraints. Vision-language-action (VLA) models provide broad generalization but introduce high inference latency and GPU memory costs, whereas vision-action chunking policies are better suited to real-time industrial control. However, such policies are typically trained with behavior cloning, which suffers from distribution shift in contact-rich tasks.
This paper proposes PAC-Act, a reinforcement learning post-training framework for pretrained action chunking Transformer policies. Key contributions:
- Reformulates policy optimization at the chunk level
- Constructs an actor-critic architecture built around ACT-style transitions
- Introduces a hybrid behavior prior constraint to anchor the fine-tuned policy to its pretrained behavior
On industrial precision contact benchmarks, PAC-Act improves task success rate, contact stability, and force safety, while maintaining low inference latency and low GPU memory usage.
---
*Auto-collected on 2026-07-14*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178395125