{"paper":{"title":"FrameSkip: Learning from Fewer but More Informative Frames in VLA Training","license":"http://creativecommons.org/licenses/by/4.0/","headline":"FrameSkip improves VLA success rates by training only on high-importance frames from demonstrations.","cross_cats":[],"primary_cat":"cs.RO","authors_text":"Bailing Wang, Bin Yu, Changti Wu, Cong Huang, Haishan Liu, Hang Yuan, Kai Chen, Shijie Lian, Xiaopeng Lin, Yuliang Wei, Zhaolong Shen","submitted_at":"2026-05-13T16:38:05Z","abstract_excerpt":"Vision-Language-Action (VLA) policies are commonly trained from dense robot demonstration trajectories, often collected through teleoperation, by sampling every recorded frame as if it provided equally useful supervision. We argue that this convention creates a temporal supervision imbalance: long low-change segments dominate the training stream, while manipulation-critical transitions such as alignment, contact, grasping, and release appear only sparsely. We introduce FrameSkip, a data-layer frame selection framework that scores trajectory frames using action variation, visual-action coherenc"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across RoboCasa-GR1, SimplerEnv, and LIBERO, FrameSkip improves the success-retention trade-off over full-frame training and simpler frame selection variants, achieving a macro-average success rate of 76.15% across the three benchmarks compared with 66.50% for full-frame training while using a compressed trajectory view that retains 20% of unique frames in the main setting.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the four scoring signals (action variation, visual-action coherence, task-progress priors, gripper-transition preservation) reliably identify manipulation-critical frames without systematic bias or omission of important transitions on unseen tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"FrameSkip improves VLA policy training success from 66.50% to 76.15% by selecting high-importance frames and retaining only 20% of unique frames across three benchmarks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"FrameSkip improves VLA success rates by training only on high-importance frames from demonstrations.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"99386c3f9e4b9d262a7fecb0869299d2f0148dced6c3f0d7f66b3f66d557abca"},"source":{"id":"2605.13757","kind":"arxiv","version":1},"verdict":{"id":"fd3e7c37-9008-4bb6-9a1b-39ae2a3eee90","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T17:55:54.542777Z","strongest_claim":"Across RoboCasa-GR1, SimplerEnv, and LIBERO, FrameSkip improves the success-retention trade-off over full-frame training and simpler frame selection variants, achieving a macro-average success rate of 76.15% across the three benchmarks compared with 66.50% for full-frame training while using a compressed trajectory view that retains 20% of unique frames in the main setting.","one_line_summary":"FrameSkip improves VLA policy training success from 66.50% to 76.15% by selecting high-importance frames and retaining only 20% of unique frames across three benchmarks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the four scoring signals (action variation, visual-action coherence, task-progress priors, gripper-transition preservation) reliably identify manipulation-critical frames without systematic bias or omission of important transitions on unseen tasks.","pith_extraction_headline":"FrameSkip improves VLA success rates by training only on high-importance frames from demonstrations."},"references":{"count":21,"sample":[{"doi":"","year":null,"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","ref_index":1,"cited_arxiv_id":"2410.24164","is_internal_anchor":true},{"doi":"","year":null,"title":"InternVLA-M1: A Spatially Guided Vision-Language-Action Framework for Generalist Robot Policy","work_id":"8a11d29e-4bf8-4a9c-a97e-d87e7350dd9c","ref_index":2,"cited_arxiv_id":"2510.13778","is_internal_anchor":true},{"doi":"","year":null,"title":"Robot data curation with mutual information estimators","work_id":"9999f054-351c-44bd-bbf5-62ce3a762e6d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":null,"title":"Thinkact: Vision- language-action reasoning via reinforced visual latent planning","work_id":"a5c17d1f-5739-41bd-bd90-0faf9424af13","ref_index":4,"cited_arxiv_id":"2507.16815","is_internal_anchor":false},{"doi":"","year":null,"title":"$\\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization","work_id":"d1ad7304-d09a-49bc-809e-846439f6aff9","ref_index":5,"cited_arxiv_id":"2504.16054","is_internal_anchor":true}],"resolved_work":21,"snapshot_sha256":"b9cc3329d285d0b24670f775506460bb87897f583ce871eae3424922b3b3d1d8","internal_anchors":14},"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"}