{"paper":{"title":"What Matters in Building Vision-Language-Action Models for Generalist Robots","license":"http://creativecommons.org/licenses/by-nc-nd/4.0/","headline":"Specific choices in backbones, architectures, and data timing let simple Vision-Language-Action models set new robot manipulation records.","cross_cats":["cs.CV"],"primary_cat":"cs.RO","authors_text":"Bingyi Kang, Di Guo, Dong Wang, Hanbo Zhang, Huaping Liu, Jirong Liu, Long Qian, Minghuan Liu, Peiyan Li, Tao Kong, Xiao Ma, Xinghang Li, Xinlong Wang","submitted_at":"2024-12-18T17:07:20Z","abstract_excerpt":"To utilize Foundation Vision Language Models (VLMs) for robotic tasks and motion planning, the community has proposed different methods for injecting action components into VLMs and building the Vision-Language-Action models (VLAs). In this work, we disclose the key factors that significantly influence the performance of VLA on robot manipulation problems and focus on answering three essential design choices: which backbone to select, how to formulate the VLA architectures, and when to add cross-embodiment data. The obtained results convince us firmly to explain why we prefer VLA and develop a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Through extensive experiments which include over 8 VLM backbones, 4 policy architectures, and over 600 distinct designed experiments, we provide a detailed guidebook for the future design of VLAs... RoboVLMs... achieve a new state-of-the-art performance in three simulation tasks and real-world experiments.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the chosen simulation tasks and real-world experiments are representative enough of generalist robot manipulation that the observed ranking of design choices will transfer to new tasks, embodiments, and environments not tested in the 600+ runs.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Systematic tests of VLM backbones, policy architectures, and cross-embodiment data yield RoboVLMs that set new SOTA on robot manipulation benchmarks while requiring few manual designs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Specific choices in backbones, architectures, and data timing let simple Vision-Language-Action models set new robot manipulation records.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"233bf637abbb829875db7cf18e555ed55f389674e64b32d2597d2ef338c8487f"},"source":{"id":"2412.14058","kind":"arxiv","version":4},"verdict":{"id":"5de490e9-804b-4794-af7a-10c5bab1f587","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T21:33:48.446845Z","strongest_claim":"Through extensive experiments which include over 8 VLM backbones, 4 policy architectures, and over 600 distinct designed experiments, we provide a detailed guidebook for the future design of VLAs... RoboVLMs... achieve a new state-of-the-art performance in three simulation tasks and real-world experiments.","one_line_summary":"Systematic tests of VLM backbones, policy architectures, and cross-embodiment data yield RoboVLMs that set new SOTA on robot manipulation benchmarks while requiring few manual designs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the chosen simulation tasks and real-world experiments are representative enough of generalist robot manipulation that the observed ranking of design choices will transfer to new tasks, embodiments, and environments not tested in the 600+ runs.","pith_extraction_headline":"Specific choices in backbones, architectures, and data timing let simple Vision-Language-Action models set new robot manipulation records."},"references":{"count":54,"sample":[{"doi":"","year":2022,"title":"Flamingo: a visual language model for few-shot learning","work_id":"abe0dc9d-38b0-4f78-87e4-1f56476c29c8","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond","work_id":"cbc2bb21-b6bb-46c0-80bf-107e195ffe10","ref_index":2,"cited_arxiv_id":"2308.12966","is_internal_anchor":true},{"doi":"","year":2024,"title":"PaliGemma: A versatile 3B VLM for transfer","work_id":"df6f48b3-5792-47c7-9614-cb856ea31ad9","ref_index":3,"cited_arxiv_id":"2407.07726","is_internal_anchor":true},{"doi":"","year":2024,"title":"$\\pi_0$: A Vision-Language-Action Flow Model for General Robot Control","work_id":"f790abdc-a796-482f-a40d-f8ee035ecfc2","ref_index":4,"cited_arxiv_id":"2410.24164","is_internal_anchor":true},{"doi":"","year":2023,"title":"RoboCat : A self-improving foundation agent for robotic manipulation","work_id":"143e7731-0488-4088-8e23-63f9d4140118","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":54,"snapshot_sha256":"068783adc0418b56d569a8d0f13ce528ebe66b651ad9e79116787bc514b71dec","internal_anchors":25},"formal_canon":{"evidence_count":2,"snapshot_sha256":"8df15057d85122e96a54c885c04271576961ff57df3c998dfacce45ccdae174c"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}