Research map / Reinforcement learning / RL from feedback
SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning
Haozhan Li, Yuxin Zuo, Jiale Yu, Yuhao Zhang, Yang Zhaohui, Kaiyan Zhang, Xuekai Zhu, Yuchen Zhang, Tianxing Chen, Ganqu Cui, Dehui Wang, Dingxiang Luo and 9 more
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 SimpleVLA-RL on top of full-data SFT | OpenVLA-OFT · LIBERO average (Full-Trajectory SFT) | success rate (%): 91.0 -> 99.1 | Improved |
| apply SimpleVLA-RL on top of one-demonstration-per-task SFT | OpenVLA-OFT · LIBERO-Long (One-Trajectory SFT) | success rate (%): 17.3 -> 91.7 | Improved |
| apply SimpleVLA-RL on top of SFT for dual-arm tasks | OpenVLA-OFT · RoboTwin2.0 (12 tasks overall average) | success rate (%): 38.3 -> 68.8 | Improved |
| apply SimpleVLA-RL on SFT model trained only in simulation, evaluated on real robot | OpenVLA-OFT · Real-world sim2real (4 tasks average) | success rate (%): 17.5 -> 38.5 | Improved |
| use RL instead of SFT on 9 seen tasks, evaluating on held-out unseen tasks (ablation) | OpenVLA-OFT (One-Trajectory SFT base) · LIBERO Spatial/Object/Goal unseen tasks | success rate on unseen task: Spatial Unseen Task 1: 43.3 -> 71.8 | Improved |
| add exploration enhancements: dynamic sampling, clip higher (0.28), temperature 1.6 (ablation) | SimpleVLA-RL · LIBERO | success rate: 10-15% consistent improvement | Improved |
| apply RL to a base model with no SFT trajectories (ablation) | OpenVLA-OFT · RoboTwin2.0 5 tasks, 0 trajectories SFT | success rate (%): 0 -> 0 | No clear effect |
| increase SFT trajectories from 100 to 1000 before RL (ablation) | SimpleVLA-RL · RoboTwin2.0 5 tasks average | success rate after RL (%): 25.4 -> 50.4 | Improved |
| apply RL when initial success rate is very low (100 trajectories SFT) (ablation) | SimpleVLA-RL · RoboTwin2.0 Pick Dual Bottles | success rate (%): 1.2 -> 4.3 (vs 29.7 -> 68.3 with 1000 trajectories) | Mixed |
| remove KL regularization and reference model | SimpleVLA-RL · training | memory and exploration | Improved |
Later papers that compared with this method
1 later accepted paper reports SimpleVLA-RL as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- DyGRO-VLA: Cross-Task Scaling of Vision–Language–Action Models via Dynamic Grouped Residual Optimization (ICML 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- NS-VLA: Towards Neuro-Symbolic Vision-Language-Action Models (NeurIPS 2026 · similarity 0.93)
- ElegantVLA: Learning When to Think for Efficient Vision-Language-Action Models (NeurIPS 2026 · similarity 0.92)
- What Can RL Bring to VLA Generalization? An Empirical Study (NeurIPS 2025 · similarity 0.92)
- Beyond Data Scaling: Representation-Centric Pre-training for Vision-Language-Action Models (NeurIPS 2026 · similarity 0.91)
- VLA-REPLICA: A Low-Cost, Reproducible Benchmark for Real-World Evaluation of Vision-Language-Action Models (NeurIPS 2026 · similarity 0.90)
- Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes (AAAI 2025 · similarity 0.90)
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.
