{"paper":{"title":"HybridVLA: Collaborative Diffusion and Autoregression in a Unified Vision-Language-Action Model","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"HybridVLA unifies diffusion for continuous actions and autoregression for reasoning inside one vision-language-action model.","cross_cats":["cs.RO"],"primary_cat":"cs.CV","authors_text":"Chengkai Hou, Chenyang Gu, Hao Chen, Jiaming Liu, KC alex Zhou, Mengdi Zhao, Mengzhen Liu, Pengju An, Pheng-Ann Heng, Renrui Zhang, Shanghang Zhang, Sixiang Chen, Xiaoqi Li, Zhuoyang Liu, Ziyu Guo","submitted_at":"2025-03-13T17:59:52Z","abstract_excerpt":"A fundamental objective of manipulation policy design is to endow robots to comprehend human instructions, reason about scene cues, and execute generalized actions in dynamic environments. Recent autoregressive vision-language-action (VLA) methods inherit common-sense reasoning capabilities from vision-language models (VLMs) for next action-token prediction. However, these methods quantize actions into discrete bins, which disrupts the continuity required for precise control. In contrast, existing diffusion-based VLA methods incorporate an additional diffusion head to predict continuous action"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"HybridVLA outperforms previous state-of-the-art VLA methods by 14% and 19% in mean success rate on simulation and real-world tasks, respectively, while demonstrating stable manipulation in unseen configurations.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The collaborative training recipe successfully prevents interference between diffusion denoising and next-token prediction while allowing the two paradigms to reinforce each other across tasks.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"HybridVLA unifies diffusion and autoregression in a single VLA model via collaborative training and ensemble to raise robot manipulation success rates by 14% in simulation and 19% in real-world tasks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"HybridVLA unifies diffusion for continuous actions and autoregression for reasoning inside one vision-language-action model.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"f60d62d555cb87ba1712b6516e4b2f2529fa7f4897ada1c2047204adbd6a44dd"},"source":{"id":"2503.10631","kind":"arxiv","version":3},"verdict":{"id":"17f48075-055a-45b6-b90c-25bba68fbb29","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T21:57:10.964919Z","strongest_claim":"HybridVLA outperforms previous state-of-the-art VLA methods by 14% and 19% in mean success rate on simulation and real-world tasks, respectively, while demonstrating stable manipulation in unseen configurations.","one_line_summary":"HybridVLA unifies diffusion and autoregression in a single VLA model via collaborative training and ensemble to raise robot manipulation success rates by 14% in simulation and 19% in real-world tasks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The collaborative training recipe successfully prevents interference between diffusion denoising and next-token prediction while allowing the two paradigms to reinforce each other across tasks.","pith_extraction_headline":"HybridVLA unifies diffusion for continuous actions and autoregression for reasoning inside one vision-language-action model."},"references":{"count":127,"sample":[{"doi":"","year":2023,"title":"PaLM-E: An Embodied Multimodal Language Model","work_id":"5b99811a-1d93-47e2-9d59-f4045a0b74a2","ref_index":1,"cited_arxiv_id":"2303.03378","is_internal_anchor":true},{"doi":"","year":2022,"title":"Do As I Can, Not As I Say: Grounding Language in Robotic Affordances","work_id":"037320f1-b0a9-4cbe-a639-bfb25409ce71","ref_index":2,"cited_arxiv_id":"2204.01691","is_internal_anchor":true},{"doi":"","year":2023,"title":"VoxPoser: Composable 3D Value Maps for Robotic Manipulation with Language Models","work_id":"5a5edf95-2538-4e2b-8dfa-da39cec89f22","ref_index":3,"cited_arxiv_id":"2307.05973","is_internal_anchor":true},{"doi":"","year":2024,"title":"What Matters in Building Vision-Language-Action Models for Generalist Robots","work_id":"972ba6c7-a312-404e-bd8d-851850b28fc9","ref_index":4,"cited_arxiv_id":"2412.14058","is_internal_anchor":true},{"doi":"","year":2024,"title":"Visual instruction tuning","work_id":"a2628f23-eb28-4e09-a098-bce23f22dee1","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":127,"snapshot_sha256":"6828e62d3c65ce92fbd7fb78972eb241f48387976896d9433f37739a35edf8b1","internal_anchors":42},"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"}