VLAFlow: A Unified Training Framework for Vision-Language-Action Models via Co-training and Future Latent Alignment

2026-07-01

AI Summary

This work addresses the challenge of fairly comparing different pretraining paradigms for vision-language-action (VLA) models, which has been hindered by inconsistent architectures, datasets, and evaluation protocols. To this end, the authors propose VLAFlow, a unified framework that systematically evaluates four pretraining paradigms under a shared vision-language backbone, a 14-dimensional meta-action space, and the heterogeneous robot dataset OXEMix. Through co-training and future latent-space alignment, the study reveals—for the first time—the complementary nature of language supervision and latent-space alignment. The introduced meta-action space enhances the smoothness and transferability of heterogeneous action supervision. The proposed MindLWPI composite method achieves the most stable cross-benchmark transfer performance on LIBERO, LIBERO-Plus, and SimplerEnv, significantly outperforming any single pretraining paradigm.

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.

Problem

• vision-language-action models
• pre-training paradigms
• heterogeneous robot data
• controlled comparison
• action space

Innovation

• vision-language-action models
• co-training
• future latent alignment
• flow-matching framework
• heterogeneous robot data