Research map / Vision-language and multimodal / Robot action and navigation
PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models
Xianghui Wang, Feng Chen, Wenbo Zhang, Hua Yan, Zixuan Wang, Changsheng Li, Yinjie Lei
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply PolicyTrim (two-stage RL post-training) to the VLA | pi0.5 · LIBERO-Object | physical steps: 125.0 -> 64.3 | Improved |
| apply PolicyTrim | pi0.5 · LIBERO-Object | success rate: 99.1 -> 98.5 | No clear effect |
| add reliable chunk extension alone (ablation) | pi0.5 · LIBERO-Spatial | physical steps: 108.3 -> 113.8 | Hurt |
| add reliable chunk extension alone (ablation) | pi0.5 · LIBERO-Spatial | end-to-end speedup: 1.0 -> 2.86 | Improved |
| add step-saving reward alone (no regularization) (ablation) | pi0.5 · LIBERO-Spatial | success rate: 97.8 -> 93.7 | Hurt |
| add step-saving reward alone (ablation) | pi0.5 · LIBERO-Spatial | physical steps: 108.3 -> 81.7 | Improved |
| add group-anchored regularization to the step-saving reward (ablation) | pi0.5 · LIBERO-Spatial | success rate: 93.7 -> 97.5 | Improved |
| add group-anchored regularization to the step-saving reward (ablation) | pi0.5 · LIBERO-Spatial | physical steps: 81.7 -> 61.6 | Improved |
| use dynamic horizon exploration instead of fixed gamma=1.0 (ablation) | pi0.5 · LIBERO-Object (H=20) | success rate: 94.4 -> 98.8 | Improved |
| increase fixed acceptance ratio from 0.25 to 1.0 (ablation) | pi0.5 · LIBERO-Object (H=20) | success rate: 99.2 -> 94.4 | Hurt |
| increase GRPO group size from 8 to 16 (ablation) | PolicyTrim pi0.5 · LIBERO-Spatial | success rate: 97.8 -> 97.3 | No clear effect |
| apply PolicyTrim step reduction only to OpenVLA-OFT | OpenVLA-OFT · LIBERO-Spatial | end-to-end speedup: 1.0 -> 1.79 | Improved |
Closest papers by meaning
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon (NeurIPS 2026 · similarity 0.91)
- RePO-VLA: Recovery-Driven Policy Optimization for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching (NeurIPS 2025 · similarity 0.91)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
