{"ID":5676063,"CreatedAt":"2026-07-03T01:40:09.565152011Z","UpdatedAt":"2026-07-05T01:25:09.323207391Z","DeletedAt":null,"paper_url":"https://arxiv.org/abs/2607.01586","arxiv_id":"2607.01586","title":"VLAFlow: A Unified Training Framework for Vision-Language-Action Models via Co-training and Future Latent Alignment","abstract":"Vision-language-action models (VLAs) have recently advanced robotic manipulation, yet the effects of different robot-data pre-training paradigms remain difficult to compare because existing models often differ in architecture, data, action space, and evaluation protocol. We present VLAFlow (Vision-Language-Action Flow), a unified flow-matching framework for controlled comparison of VLA training objectives. Using a heterogeneous robot corpus, OXEMix, containing approximately 5,000 hours of data from DROID, OpenX-Embodiment, OpenX-Augmented, and RoboCOIN, we evaluate four paradigms under the same pi0-style architecture, shared VLM backbone, action expert, and 14-dimensional action space: action-only modeling (MindPI), language-supervised co-training (MindLPI), future latent alignment (MindWPI), and their combination (MindLWPI). Experiments on LIBERO, LIBERO-Plus, and SimplerEnv show that action-only pre-training is sensitive to heterogeneous data. In contrast, language supervision helps preserve vision-language generalization, while future latent alignment improves state-transition and action-outcome modeling. By combining both signals, MindLWPI achieves the most stable overall transfer performance across benchmarks. These results suggest a meta-action space view: language and future latent representations provide complementary intermediate constraints that make heterogeneous action supervision smoother and more transferable.","short_abstract":"Vision-language-action models (VLAs) have recently advanced robotic manipulation, yet the effects of different robot-data pre-training paradigms remain difficult to compare because existing models often differ in architecture, data, action space, and evaluation protocol. We present VLAFlow (Vision-Language-Action Flow)...","url_abs":"https://arxiv.org/abs/2607.01586","url_pdf":"https://arxiv.org/pdf/2607.01586v1","authors":"[\"Guoyang Xia\",\"Fengfa Li\",\"Hongjin Ji\",\"Lei Ren\",\"Fangxiang Feng\",\"Kun Zhan\",\"Yan Xie\"]","published":"2026-07-02T01:38:16Z","proceeding":"cs.CV","tasks":"[\"cs.CV\",\"cs.AI\",\"cs.RO\"]","methods":"[]","has_code":false}