{"paper":{"title":"TinyVLA: Towards Fast, Data-Efficient Vision-Language-Action Models for Robotic Manipulation","license":"http://creativecommons.org/licenses/by/4.0/","headline":"TinyVLA reaches OpenVLA-level performance on robot tasks by initializing from fast multimodal models and adding a diffusion action decoder, removing the pre-training stage entirely.","cross_cats":["cs.CV"],"primary_cat":"cs.RO","authors_text":"Chaomin Shen, Feifei Feng, Jian Tang, Jinming Li, Junjie Wen, Kun Wu, Minjie Zhu, Ning Liu, Ran Cheng, Yaxin Peng, Yichen Zhu, Zhiyuan Xu","submitted_at":"2024-09-19T07:10:18Z","abstract_excerpt":"Vision-Language-Action (VLA) models have shown remarkable potential in visuomotor control and instruction comprehension through end-to-end learning processes. However, current VLA models face significant challenges: they are slow during inference and require extensive pre-training on large amounts of robotic data, making real-world deployment difficult. In this paper, we introduce a new family of compact vision-language-action models, called TinyVLA, which offers two key advantages over existing VLA models: (1) faster inference speeds, and (2) improved data efficiency, eliminating the need for"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Our approach significantly outperforms the state-of-the-art VLA model, OpenVLA, in terms of speed and data efficiency, while delivering comparable or superior performance.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That initializing the policy backbone with existing high-speed multimodal models plus a diffusion decoder during fine-tuning is sufficient to eliminate the pre-training stage while preserving or improving task performance and generalization.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"TinyVLA achieves faster inference and higher data efficiency than OpenVLA on robotic manipulation tasks by initializing from high-speed multimodal models and adding a diffusion policy decoder, without any pre-training phase.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"TinyVLA reaches OpenVLA-level performance on robot tasks by initializing from fast multimodal models and adding a diffusion action decoder, removing the pre-training stage entirely.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d2290bf4b1530764ac0538952466e4728758aebbbdb0d343407da3c0cedc10b3"},"source":{"id":"2409.12514","kind":"arxiv","version":5},"verdict":{"id":"b4940e26-1ddf-4bd7-8052-f922074b9249","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-17T16:07:08.380660Z","strongest_claim":"Our approach significantly outperforms the state-of-the-art VLA model, OpenVLA, in terms of speed and data efficiency, while delivering comparable or superior performance.","one_line_summary":"TinyVLA achieves faster inference and higher data efficiency than OpenVLA on robotic manipulation tasks by initializing from high-speed multimodal models and adding a diffusion policy decoder, without any pre-training phase.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That initializing the policy backbone with existing high-speed multimodal models plus a diffusion decoder during fine-tuning is sufficient to eliminate the pre-training stage while preserving or improving task performance and generalization.","pith_extraction_headline":"TinyVLA reaches OpenVLA-level performance on robot tasks by initializing from fast multimodal models and adding a diffusion action decoder, removing the pre-training stage entirely."},"references":{"count":46,"sample":[{"doi":"","year":2024,"title":"Roboagent: Generalization and efficiency in robot manipulation via semantic augmentations and action chunking,","work_id":"ad2bd81c-bdcd-4b21-bc80-5819aaa4bea7","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Bridge data: Boosting generalization of robotic skills with cross-domain datasets,","work_id":"fdece732-d4e5-450c-8f75-6ca0dbdf84ac","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Diffusion policy: Visuomotor policy learning via action diffusion,","work_id":"eaca3997-9409-4ca0-a030-e58e1eae24f6","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"3d diffusion policy: Generalizable visuomo- tor policy learning via simple 3d representations,","work_id":"709db3be-7760-4bb5-bad0-1fc5c10cfd2f","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2023,"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","work_id":"68a5177f-d644-44c1-bd4f-4e5278c22f5d","ref_index":5,"cited_arxiv_id":"2307.09288","is_internal_anchor":true}],"resolved_work":46,"snapshot_sha256":"40a7bc759fc5fcf4399418d23e48a0662db7903bbb5d67e27a2eacbe096ef5bc","internal_anchors":11},"formal_canon":{"evidence_count":1,"snapshot_sha256":"03095fa8158cd570a3a929796d717e55d443cc61cc3ce48bb56e3987742f6747"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}