PAC-ACT

Post-training Actor-Critic for Action Chunking Transformers

Yujie Pang, Zudong Li

arXiv preprint · July 2026

PAC-ACT architecture · From observations and the pretrained ACT policy to actor–critic updates.
PAC-ACT architecture · From observations and the pretrained ACT policy to actor–critic updates.

Overview

A contact policy can reach the right area and still apply too much force. PAC-ACT starts from a pretrained Action Chunking Transformer and fine-tunes how it acts when the starting pose and contact conditions change.

Learning under contact

The actor and critic operate on action chunks. A behavior prior constrains the update so that the policy can explore while retaining the action distribution learned from demonstrations. On the Contour task, the paper reports a 46-fold reduction in the proportion of force readings above 60 N.

Contact experiment

The video shows an ACT baseline episode in MuJoCo. The paper contains the comparisons with reinforcement-learning post-training.

Paper

Open the full paper PDF

PAC-ACT: Post-training Actor-Critic for Action Chunking Transformers

Yujie Pang, Zudong Li
arXiv preprint · July 2026

BibTeX

@misc{pang2026pacact,
  title = {PAC-ACT: Post-training Actor-Critic for Action Chunking Transformers},
  author = {Pang, Yujie and Li, Zudong},
  year = {2026},
  eprint = {2607.09590},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2607.09590}
}
Download .bib