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

Dynin-Robotics. Omnimodal Unified Diffusion Vision-Language-Action Model. By Hoeun Lee and six others, from Seoul National University. Published on the eleventh of September. Its code repository is a placeholder: a readme, and nothing to run yet.

A language-conditioned robot policy turns an instruction and what it sees into actions. Two kinds of prediction could help it. A visual goal gives it a target outcome. Dynamics prediction gives it a representation of how the scene changes depending on the action taken. The question the authors take up is how to bring both predictions into action generation and action selection themselves. Predicting a goal or the next frame is one thing. Using those predictions when choosing what the robot does next is another, and that is the step this work is about.

Dynin-Robotics answers with a shared trajectory model. It is built on Dynin- Omni, an omnimodal masked-diffusion backbone, and it represents language, visual observations, goals and actions all as discrete tokens. By varying which spans serve as conditioning and which serve as targets, one model learns four jobs: action prediction, next-observation prediction given an action, terminal goal-state prediction, and reconstruction of the instruction from a trajectory. These are not separate networks. They are the same model, asked different questions by changing which parts of a trajectory it is shown and which parts it must fill in.

Those interfaces support test-time scaling in three ways: through goal prediction, through evaluating candidate actions, and through jointly refining actions and predictions of future states. The model is continually pretrained on about one point three three million trajectories drawn from forty-eight Open X-Embodiment datasets, and then adapted separately to each downstream domain. Those extra predictions are not only training signals. At test time they become tools: predict the goal, score several candidate actions against it, and refine an action together with the future state it should produce.

On two VLABench tasks, robot pretraining improves adaptation within a fixed Stage-2 step budget. The full mixture of objectives improves success on shifted instructions over Policy-only post-training, under the same coupled decoder. Combining goal guidance with joint denoising of action and next state improves shifted-instruction success further, compared with action-only decoding, though the benefit depends on how the predictions are composed. Shifted instructions are where both the extra objectives and the combined decoding are reported to help, each measured against a simpler alternative.

On LIBERO and zero-shot LIBERO-Plus, performance is competitive. On a Franka Research 3 robot, the average success rate across four manipulation conditions is seventy-eight point four percent. And an optimized block-parallel implementation accelerates model-side action decoding by up to twenty-nine point two times relative to the base implementation, under the profiling setup the authors report.

The case being made is for one trajectory model as a common interface. Goals, dynamics, actions and instructions are learned as complementary objectives in the same token space, and their predictions can then be composed while the robot is being controlled, rather than each living in a separate model.