Research map / Vision-language and multimodal / Robot action and navigation
PARL-VLA: Pruning-Aware On-Policy Reinforcement Learning for Vision-Language-Action Model
ruiyan xu, Jiashu Lv, Sixu Lin, Ruixing Jin, Shuliang He, Guiliang Liu
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 |
|---|---|---|---|
| train with pruning inside the RL loop (PARL-VLA High) rather than full-token RLinf | OpenVLA-OFT + GRPO · LIBERO-Plus mean | success rate (%): 70.2 -> 74.0 | Improved |
| train with pruning in the RL loop rather than pruning only at evaluation (FastV) | OpenVLA-OFT + GRPO · LIBERO-Plus mean | success rate (%): 67.0 -> 74.0 | Improved |
| prune only at evaluation (VLA-Pruner) on a full-token-trained policy | RLinf · LIBERO-Plus mean | success rate (%): 70.2 -> 66.5 | Hurt |
| add PARL-VLA pruning-aware training | RLinf · LIBERO-Plus Light | success rate (%): 62.7 -> 68.7 | Improved |
| add PARL-VLA pruning-aware training | RLinf · LIBERO-Plus Goal suite mean | success rate (%): 66.7 -> 73.5 | Improved |
| add PARL-VLA pruning-aware training | RLinf · LIBERO in-distribution avg | success rate (%): 94.4 -> 94.9 | No clear effect |
| evaluation-only pruning (FastV) instead of full tokens | RLinf · LIBERO in-distribution avg | success rate (%): 94.4 -> 92.4 | Hurt |
| replay the recorded rollout pruning configuration during actor updates (vs full tokens at update) (ablation) | PARL-VLA · LIBERO training stability | approx. KL: KL near 0.01 vs above 1 | Improved |
| replay the rollout pruning configuration rather than using a different update budget (min(1,1.1k)) (ablation) | PARL-VLA · LIBERO training | success rate | Improved |
| add spectrum training over keep ratios vs fixed rho=0.40 (ablation) | PARL-VLA · LIBERO-Plus avg | success rate (%): 71.1 -> 73.5 | Improved |
| add spectrum training vs fixed rho=0.40 (ablation) | PARL-VLA · LIBERO-Plus worst-case | success rate (%): 69.2 -> 72.7 | Improved |
| add spectrum training vs fixed rho=0.30 at eval keep ratio 0.30 (ablation) | PARL-VLA · LIBERO-Plus eval rho=0.30 | success rate (%): 63.9 -> 72.7 | Improved |
Closest papers by meaning
- What Can RL Bring to VLA Generalization? An Empirical Study (NeurIPS 2025 · similarity 0.90)
- When and What to Prune? Stage-Aware Visual Token Pruning for Efficient VLA (NeurIPS 2026 · similarity 0.90)
- NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning (CVPR 2026 · similarity 0.89)
- EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models (NeurIPS 2025 · similarity 0.89)
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.89)
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.
