| --- |
| license: apache-2.0 |
| language: |
| - en |
| library_name: pytorch |
| tags: |
| - robotics |
| - vision-language-action |
| - reinforcement-learning |
| - grpo |
| - policy-efficiency |
| - embodied-ai |
| - libero |
| - maniskill |
| - metaworld |
| - openpi |
| - openvla |
| arxiv: 2606.22540 |
| --- |
| |
| <div align="center"> |
|
|
| # PolicyTrim |
|
|
| ### Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models |
|
|
| [](https://arxiv.org/abs/2606.22540) |
| [](https://github.com/INCEPTIONwang/PolicyTrim) |
| [](https://inceptionwang.github.io/PolicyTrim/) |
| [](https://huggingface.co/papers/2606.22540) |
| [](https://github.com/INCEPTIONwang/PolicyTrim/blob/main/LICENSE) |
|
|
| **Xianghui Wang\***, **Feng Chen\***, Wenbo Zhang, Hua Yan, Zixuan Wang<sup>†</sup>, Changsheng Li, Yinjie Lei<sup>‡</sup> |
|
|
| <sup>*</sup> Equal contribution · <sup>†</sup> Project lead · <sup>‡</sup> Corresponding author |
| |
| </div> |
| |
| ## Model Card |
| |
| This repository provides the released post-training actor checkpoints for |
| **PolicyTrim**, a two-stage reinforcement learning framework for improving the |
| intrinsic policy efficiency of Vision-Language-Action (VLA) models. |
| |
| Most deployment-efficiency methods reduce the latency of each model forward |
| pass. PolicyTrim instead reduces how many inference calls and physical actions |
| are required to finish a task. It targets two policy-level bottlenecks: |
| |
| 1. unreliable predictions near the tail of an action chunk; |
| 2. redundant physical execution steps and corrective actions. |
| |
| PolicyTrim first extends the reliable executable action horizon, then applies a |
| redundancy-aware step-saving objective with stability regularization. Across |
| three benchmarks and three VLA model families, the method reports: |
| |
| - **3x** improvement in action chunk utilization; |
| - **51.4%** reduction in physical execution steps; |
| - up to **5.83x** end-to-end deployment speedup; |
| - no compromise in task success rates. |
| |
| For the method, training code, configuration files, and evaluation scripts, see |
| the [PolicyTrim GitHub repository](https://github.com/INCEPTIONwang/PolicyTrim). |
| |
| <div align="center"> |
| <img src="https://raw.githubusercontent.com/INCEPTIONwang/PolicyTrim/main/overview_01.png" alt="PolicyTrim overview" width="100%"/> |
| </div> |
| |
| ## Resources |
| |
| - **Paper:** [arXiv:2606.22540](https://arxiv.org/abs/2606.22540) |
| - **PDF:** [PolicyTrim paper](https://arxiv.org/pdf/2606.22540) |
| - **Project page:** [inceptionwang.github.io/PolicyTrim](https://inceptionwang.github.io/PolicyTrim/) |
| - **Code:** [github.com/INCEPTIONwang/PolicyTrim](https://github.com/INCEPTIONwang/PolicyTrim) |
| - **Hugging Face paper page:** [huggingface.co/papers/2606.22540](https://huggingface.co/papers/2606.22540) |
| |
| ## Download |
| |
| Install the Hugging Face Hub CLI: |
| |
| ```bash |
| pip install -U huggingface_hub |
| ``` |
| |
| Download the complete repository: |
| |
| ```bash |
| hf download INCEPTIONwang/PolicyTrim \ |
| --local-dir ./PolicyTrim-checkpoints |
| ``` |
| |
| The complete repository is large. To download only one checkpoint, specify its |
| path. For example: |
| |
| ```bash |
| hf download INCEPTIONwang/PolicyTrim \ |
| libero_goal_grpo_openpi_pi05/checkpoints/global_step_500/actor/model_state_dict/full_weights.pt \ |
| --local-dir ./PolicyTrim-checkpoints |
| ``` |
| |
| Python equivalent: |
| |
| ```python |
| from huggingface_hub import hf_hub_download |
| |
| checkpoint_path = hf_hub_download( |
| repo_id="INCEPTIONwang/PolicyTrim", |
| filename=( |
| "libero_goal_grpo_openpi_pi05/checkpoints/global_step_500/" |
| "actor/model_state_dict/full_weights.pt" |
| ), |
| ) |
| ``` |
| |
| ## Loading and Evaluation |
| |
| Checkpoint restoration depends on the matching VLA backend and distributed |
| training configuration. Follow the setup and evaluation instructions in the |
| [GitHub README](https://github.com/INCEPTIONwang/PolicyTrim#installation), then |
| point the corresponding PolicyTrim configuration to the downloaded checkpoint. |
| |
| ## License |
| |
| The released materials are provided under the |
| [Apache License 2.0](https://github.com/INCEPTIONwang/PolicyTrim/blob/main/LICENSE). |
| Users are also responsible for complying with the licenses and terms of the |
| corresponding base VLA models, datasets, and simulation environments. |
| |
| ## Citation |
| |
| If you find PolicyTrim useful, please cite: |
| |
| ```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} |
| } |
| ``` |
| |
