README.md

---
license: apache-2.0
pipeline_tag: robotics
---

# PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models

This repository contains the models and codebase for **PolicyTrim**, a reinforcement learning-based post-training framework that extends the reliable action chunk length and reduces redundant physical steps for Vision-Language-Action (VLA) models.

- **Paper:** [PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models](https://huggingface.co/papers/2606.22540)
- **Project Page:** [https://inceptionwang.github.io/PolicyTrim/](https://inceptionwang.github.io/PolicyTrim/)
- **Repository:** [https://github.com/INCEPTIONwang/PolicyTrim](https://github.com/INCEPTIONwang/PolicyTrim)

## Introduction

PolicyTrim is a two-stage RL post-training framework designed to reduce the number of inference calls required to complete robotic manipulation tasks. It targets two key factors:
1. **Reliable action chunk extension:** Progressively probes longer execution windows and rewards successful rollouts that sustain longer reliable chunks.
2. **Redundancy-aware step reduction:** Encourages successful trajectories that reach the goal in fewer physical steps while discouraging fragile, non-reproducible shortcuts.

Across three benchmarks (including LIBERO and ManiSkill) and three VLA models (OpenPI, OpenVLA-OFT, and GR00T), PolicyTrim improves action chunk utilization by **3×**, reduces physical execution steps by **51.4%**, and achieves up to a **5.83×** end-to-end deployment speedup without sacrificing task success rates.

## Citation

If you find PolicyTrim helpful, please cite the following work:

```bibtex
@inproceedings{policytrim2026,
  title     = {PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models},
  author    = {Xianghui Wang and Feng Chen and Wenbo Zhang and Hua Yan and Zixuan Wang and Changsheng Li and Yinjie Lei},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```