Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model

Introduction

Unified masked-diffusion modeling across policy generation, world modeling, goal-state prediction, and task understanding.

Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model is a robot foundation model built on Dynin-Omni. It represents language, visual observations, goal states, and robot actions as discrete tokens in a shared trajectory sequence, connecting semantic grounding, visual prediction, and control through one bidirectional masked-diffusion backbone.

By changing the visible context, objective token, and masked target spans, the same Transformer and prediction head generate action chunks, next visual states, terminal goal states, or task instructions. Predicted goals and visual dynamics can also participate directly in action generation and candidate selection, without separate policy, goal, or world-model networks.

Training proceeds in two stages: (1) continual robot pretraining on 1.3 million trajectories from 48 Open X-Embodiment datasets, followed by (2) domain-specific post-training. The paper reports 98.1% average success on LIBERO, 73.0% on zero-shot LIBERO-Plus, and 78.4% across four real-world manipulation conditions on a Franka Research 3 robot. An optimized block-parallel decoder improves model-side action-token throughput by up to 29.2× over the base implementation under the reported profiling setup.

Notifications

Code and model will be released soon.

Citation

@article{aidaslab2026dyninrobotics,
  title={Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model},
  author={Lee, Hoeun and Kim, Jaeik and Oh, Jusang and Kim, Jinhyeok and Choi, Geon and Kim, Hyeonggeun and Do, Jaeyoung},
  journal={arXiv preprint arXiv:2609.13053},
  year={2026}
}