{"paper":{"title":"FlashSAC: Fast and Stable Off-Policy Reinforcement Learning for High-Dimensional Robot Control","license":"http://creativecommons.org/licenses/by/4.0/","headline":"FlashSAC stabilizes off-policy RL for high-dimensional robot control by cutting gradient updates and bounding norms to limit critic errors.","cross_cats":["cs.RO"],"primary_cat":"cs.LG","authors_text":"Danica Kragic, Daniel Palenicek, Donghu Kim, Florian Vogt, Hojoon Lee, I Made Aswin Nahendra, Jaegul Choo, Jan Peters, Kinam Kim, Minho Park, Sehee Min, Takuma Seno, Youngdo Lee","submitted_at":"2026-04-06T09:03:41Z","abstract_excerpt":"Reinforcement learning (RL) is a core approach for robot control when expert demonstrations are unavailable. On-policy methods such as Proximal Policy Optimization (PPO) are widely used for their stability, but their reliance on narrowly distributed on-policy data limits accurate policy evaluation in high-dimensional state and action spaces. Off-policy methods can overcome this limitation by learning from a broader state-action distribution, yet suffer from slow convergence and instability, as fitting a value function over diverse data requires many gradient updates, causing critic errors to a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across over 60 tasks in 10 simulators, FlashSAC consistently outperforms PPO and strong off-policy baselines in both final performance and training efficiency, with the largest gains on high-dimensional tasks such as dexterous manipulation. In sim-to-real humanoid locomotion, FlashSAC reduces training time from hours to minutes.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That explicitly bounding weight, feature, and gradient norms will sufficiently curb critic error accumulation in high-dimensional spaces without removing the capacity needed for accurate value estimation or policy improvement.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FlashSAC improves training speed and final performance of off-policy RL on high-dimensional robot tasks by reducing update frequency, increasing model scale, and bounding norms to limit critic error accumulation.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FlashSAC stabilizes off-policy RL for high-dimensional robot control by cutting gradient updates and bounding norms to limit critic errors.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"350de7fc0cf5087e99812e76d522f097aa206c6fc5bd272d14c2d18301905f75"},"source":{"id":"2604.04539","kind":"arxiv","version":2},"verdict":{"id":"4f1dfd44-0e81-404c-95aa-a813e0e143f9","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T17:04:49.619467Z","strongest_claim":"Across over 60 tasks in 10 simulators, FlashSAC consistently outperforms PPO and strong off-policy baselines in both final performance and training efficiency, with the largest gains on high-dimensional tasks such as dexterous manipulation. In sim-to-real humanoid locomotion, FlashSAC reduces training time from hours to minutes.","one_line_summary":"FlashSAC improves training speed and final performance of off-policy RL on high-dimensional robot tasks by reducing update frequency, increasing model scale, and bounding norms to limit critic error accumulation.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That explicitly bounding weight, feature, and gradient norms will sufficiently curb critic error accumulation in high-dimensional spaces without removing the capacity needed for accurate value estimation or policy improvement.","pith_extraction_headline":"FlashSAC stabilizes off-policy RL for high-dimensional robot control by cutting gradient updates and bounding norms to limit critic errors."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.04539/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":94,"sample":[{"doi":"","year":2023,"title":"Loss of plasticity in continual deep reinforcement learning","work_id":"7ecff814-af37-405d-a70f-9c1fb4554063","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2020,"title":"Learning dexterous in-hand manipulation.The International Journal of Robotics Research, 39(1):3–20, 2020","work_id":"d1093acb-7a4b-42e8-aa29-51fcfd4bab71","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"A brief survey of deep reinforcement learning","work_id":"3b2842ca-9c8f-4b1a-a615-22c29e63726b","ref_index":3,"cited_arxiv_id":"1708.05866","is_internal_anchor":true},{"doi":"","year":2024,"title":"Genesis: A generative and universal physics engine for robotics and beyond, December 2024","work_id":"cb89394b-6f98-44ca-8234-f0020e9373e8","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2016,"title":"Layer Normalization","work_id":"20a2d720-0046-4c7c-bcd6-327ec8143f69","ref_index":5,"cited_arxiv_id":"1607.06450","is_internal_anchor":true}],"resolved_work":94,"snapshot_sha256":"9ea9fa4581b88687296c237e979740aa6eb68ac1e53414cdc29123fcc43c1529","internal_anchors":18},"formal_canon":{"evidence_count":2,"snapshot_sha256":"585797ed27890a08802b5db6540d6dc87e1256c228071a6ff52fcf14ba011a2f"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}