{"paper":{"title":"VLA-RL: Towards Masterful and General Robotic Manipulation with Scalable Reinforcement Learning","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"VLA-RL applies online reinforcement learning to raise pretrained vision-language-action models above finetuned baselines on robot tasks.","cross_cats":["cs.AI"],"primary_cat":"cs.RO","authors_text":"Chubin Zhang, Guanxing Lu, Haonan Jiang, Wenkai Guo, Yansong Tang, Yuheng Zhou, Zifeng Gao, Ziwei Wang","submitted_at":"2025-05-24T14:42:51Z","abstract_excerpt":"Recent high-capacity vision-language-action (VLA) models have demonstrated impressive performance on a range of robotic manipulation tasks by imitating human demonstrations. However, exploiting offline data with limited visited states will cause execution failure in out-of-distribution scenarios. Intuitively, an exploration-based method that improves on online collected data at test time could address this limitation. We present VLA-RL, an algorithmic and systematic framework that leverages online reinforcement learning (RL) to improve pretrained auto-regressive VLAs in downstream tasks. Withi"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"VLA-RL enables OpenVLA-7B to surpass the strongest finetuned baseline by 4.5% on 40 challenging robotic manipulation tasks in LIBERO, and even matches the performance of advanced commercial models such as π₀-FAST.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That a vision-language model fine-tuned on pseudo reward labels from automatically extracted task segments will provide sufficiently accurate and generalizable rewards for online RL across diverse out-of-distribution scenarios.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"VLA-RL applies online RL to pretrained VLAs, yielding a 4.5% gain over strong baselines on 40 LIBERO manipulation tasks and matching commercial models like π₀-FAST.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"VLA-RL applies online reinforcement learning to raise pretrained vision-language-action models above finetuned baselines on robot tasks.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"b6ef84180e0065c11bc639260d7b1bd6818e521518951c230cfdd864f83619d7"},"source":{"id":"2505.18719","kind":"arxiv","version":1},"verdict":{"id":"2a57462a-b22a-44d3-90e0-a7abd20e77e6","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T12:52:28.368258Z","strongest_claim":"VLA-RL enables OpenVLA-7B to surpass the strongest finetuned baseline by 4.5% on 40 challenging robotic manipulation tasks in LIBERO, and even matches the performance of advanced commercial models such as π₀-FAST.","one_line_summary":"VLA-RL applies online RL to pretrained VLAs, yielding a 4.5% gain over strong baselines on 40 LIBERO manipulation tasks and matching commercial models like π₀-FAST.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That a vision-language model fine-tuned on pseudo reward labels from automatically extracted task segments will provide sufficiently accurate and generalizable rewards for online RL across diverse out-of-distribution scenarios.","pith_extraction_headline":"VLA-RL applies online reinforcement learning to raise pretrained vision-language-action models above finetuned baselines on robot tasks."},"references":{"count":93,"sample":[{"doi":"","year":2022,"title":"Proceedings of Advances in Neural Information Processing Systems (NeurIPS) 35, 28955–28971 (2022) 2, 3","work_id":"2b02057e-69a9-41d7-91b4-09a4ee4adf92","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":1910,"title":"Solving Rubik's Cube with a Robot Hand","work_id":"81bf9cee-6de8-49f6-b967-3cb853e5ba67","ref_index":2,"cited_arxiv_id":"1910.07113","is_internal_anchor":true},{"doi":"","year":2025,"title":"Univg-r1: Reasoning guided universal visual grounding with reinforcement learning","work_id":"fad4943c-f29d-4daa-a7ba-dd6f03aac5aa","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Proceedings of Advances in Neural Information Processing Systems (NeurIPS) 35, 24639– 24654 (2022) 2","work_id":"f271101d-b0ac-4f48-8199-518a4806ce5f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"In: Proceedings of International Conference on Machine Learning (ICML)","work_id":"1973ad0d-d7fd-440f-b9b7-60d76e4519de","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":93,"snapshot_sha256":"e731bc05c403e9713eb1696c70072c04f7e7a544ffe099fa2e2283171345a01a","internal_anchors":36},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}