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

A single omnimodal masked-diffusion VLA model unifies action, dynamics, and goal prediction, reaching 78.4% average success on Franka Research 3 manipulation tasks and up to 29.2x faster action decoding.

Dynin-Robotics is built on the Dynin-Omni omnimodal masked-diffusion backbone and represents language, observations, goals, and actions as discrete tokens within one shared trajectory model. A single model learns four objectives: action prediction, action-conditioned next-observation (dynamics) prediction, terminal goal-state prediction, and trajectory-to-instruction reconstruction. The shared trajectory interface enables test-time scaling via goal prediction, action-candidate evaluation, and joint action/future-state refinement. The model is continually pretrained on approximately 1.33 million trajectories from 48 Open X-Embodiment datasets. It achieves competitive results on LIBERO and zero-shot LIBERO-Plus, a 78.4% average success rate across four manipulation conditions on a Franka Research 3 robot, and up to 29.2x faster model-side action decoding from a block-parallel implementation. Both the Hugging Face daily papers listing and the arXiv (cs.AI/cs.LG/cs.CL) report state identical figures; no source disagreement.

• Single omnimodal masked-diffusion model (Dynin-Omni backbone) learns action prediction, action-conditioned next-observation prediction, terminal goal-state prediction, and trajectory-to-instruction reconstruction.
• Continually pretrained on ~1.33 million trajectories from 48 Open X-Embodiment datasets.
• 78.4% average success rate across four manipulation conditions on a Franka Research 3 robot.
• Competitive performance on LIBERO and zero-shot LIBERO-Plus benchmarks.
• Block-parallel implementation accelerates model-side action decoding by up to 29.2x.
• Enables test-time scaling through goal prediction, action-candidate evaluation, and joint action/future-state refinement.