Reproduce “PACE: Phase-Progress-Aware Credit for Long-Horizon Embodied Manipulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Chengye Song, Jiawei Zhang, Rui Song, Shengqi Wang, Xiangrong Zhang, Ziyi Wang, Huanbin Zhou, Hongzhou Wang
A framework named PACE improves post-training for Vision-Language-Action models in complex, long-horizon manipulation tasks by generating reliable step-level credit signals and adaptively utilizing them. This approach yields up to 15.3% higher success rates on real-world robot tasks and a 26.5% reduction in critic estimation error.
Key insights
- Incorporating local temporal evidence and phase-progress coordinates into the critic significantly enhances the accuracy of remaining-cost estimation, particularly around visually ambiguous critical task transitions.
- An adaptive policy distillation curriculum, switching from positive-only guidance to full advantage-conditioned distillation based on rollout success rate, provides robust policy refinement and prevents destabilization from noisy early-stage negative feedback.
- Correcting the expected remaining cost with a scalar residual, rather than modifying the full distributional shape, offers a stable and effective way to improve value estimates without increasing output dimensionality.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.