{"paper":{"title":"SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning","license":"http://creativecommons.org/licenses/by-nc-sa/4.0/","headline":"Reinforcement learning scales vision-language-action model training beyond supervised fine-tuning","cross_cats":["cs.AI","cs.CL","cs.LG"],"primary_cat":"cs.RO","authors_text":"Bowen Zhou, Dehui Wang, Dingxiang Luo, Ganqu Cui, Haozhan Li, Jiale Yu, Jiangmiao Pang, Jia Zeng, Kaiyan Zhang, Ning Ding, Shanghang Zhang, Tianxing Chen, Xuekai Zhu, Yao Mu, Youbang Sun, Yuchen Fan, Yuchen Zhang, Yuhao Zhang, Yu Wang, Yuxin Zuo, Zhaohui Yang","submitted_at":"2025-09-11T17:59:17Z","abstract_excerpt":"Vision-Language-Action (VLA) models have recently emerged as a powerful paradigm for robotic manipulation. Despite substantial progress enabled by large-scale pretraining and supervised fine-tuning (SFT), these models face two fundamental challenges: (i) the scarcity and high cost of large-scale human-operated robotic trajectories required for SFT scaling, and (ii) limited generalization to tasks involving distribution shift. Recent breakthroughs in Large Reasoning Models (LRMs) demonstrate that reinforcement learning (RL) can dramatically enhance step-by-step reasoning capabilities, raising a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"When applied to OpenVLA-OFT, SimpleVLA-RL achieves SoTA performance on LIBERO and even outperforms π₀ on RoboTwin 1.0&2.0 with the exploration-enhancing strategies we introduce. SimpleVLA-RL not only reduces dependence on large-scale data and enables robust generalization, but also remarkably surpasses SFT in real-world tasks.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the introduced VLA-specific trajectory sampling, multi-environment rendering, and exploration-enhancing strategies remain stable and effective across different base VLA models and real-world distribution shifts without extensive additional tuning or hidden failure modes.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"SimpleVLA-RL applies tailored reinforcement learning to VLA models, reaching SoTA on LIBERO, outperforming π₀ on RoboTwin, and surpassing SFT in real-world tasks while reducing data needs and identifying a 'pushcut' phenomenon.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reinforcement learning scales vision-language-action model training beyond supervised fine-tuning","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"735af64748f3db26f279dfa5878804d25a54e38ef71705dcc1acd0db9e266d18"},"source":{"id":"2509.09674","kind":"arxiv","version":1},"verdict":{"id":"45b2e080-c87d-4ae2-a567-88007d9dc30f","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T07:55:51.457256Z","strongest_claim":"When applied to OpenVLA-OFT, SimpleVLA-RL achieves SoTA performance on LIBERO and even outperforms π₀ on RoboTwin 1.0&2.0 with the exploration-enhancing strategies we introduce. SimpleVLA-RL not only reduces dependence on large-scale data and enables robust generalization, but also remarkably surpasses SFT in real-world tasks.","one_line_summary":"SimpleVLA-RL applies tailored reinforcement learning to VLA models, reaching SoTA on LIBERO, outperforming π₀ on RoboTwin, and surpassing SFT in real-world tasks while reducing data needs and identifying a 'pushcut' phenomenon.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the introduced VLA-specific trajectory sampling, multi-environment rendering, and exploration-enhancing strategies remain stable and effective across different base VLA models and real-world distribution shifts without extensive additional tuning or hidden failure modes.","pith_extraction_headline":"Reinforcement learning scales vision-language-action model training beyond supervised fine-tuning"},"references":{"count":44,"sample":[{"doi":"","year":null,"title":"OpenVLA: An Open-Source Vision-Language-Action Model","work_id":"3e7e65c5-5aed-4fe9-8414-2092bcb31cc7","ref_index":1,"cited_arxiv_id":"2406.09246","is_internal_anchor":true},{"doi":"","year":null,"title":"A Survey on Vision-Language-Action Models: An Action Tokenization Perspective","work_id":"c1359daa-6e63-4d55-9d22-1575acbd787c","ref_index":2,"cited_arxiv_id":"2507.01925","is_internal_anchor":true},{"doi":"","year":null,"title":"Roumelio- tis, and Manoj Karkee","work_id":"5f0bf2cc-1901-4ad0-940b-1e742cc6d7e7","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"What Matters in Learning from Offline Human Demonstrations for Robot Manipulation","work_id":"6a4c95c5-540e-4854-946d-c7c8a6c540ba","ref_index":4,"cited_arxiv_id":"2108.03298","is_internal_anchor":true},{"doi":"","year":null,"title":"AgiBot World Colosseo: A Large-scale Manipulation Platform for Scalable and Intelligent Embodied Systems","work_id":"f797e9ec-510f-43a7-8a0c-18009ce332e5","ref_index":5,"cited_arxiv_id":"2503.06669","is_internal_anchor":true}],"resolved_work":44,"snapshot_sha256":"d56d9d3851f4830c21737e9f637276066818a01373f7943de31e3b3af4e26396","internal_anchors":30},"formal_canon":{"evidence_count":2,"snapshot_sha256":"9612d031635a20bc73681899721228d84ce248567d78be55424283623196c2de"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}