ZeroHour
arXiv cs.AI / cs.LG / cs.CLpublished ()ingested Hoeun Lee1
Part of a story covered by 2 sources: “Dynin-Robotics: Omnimodal Unified Diffusion Vision-Language-Action Model” — merged summary and timeline →

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

infoAI researchimportance 35
AI summary · glm-5.3

Dynin-Robotics omnimodal diffusion vision-language-action model unifies action, dynamics, and goal prediction, reaching 78.4% success on Franka tasks.

Dynin-Robotics builds a shared trajectory model on the Dynin-Omni omnimodal masked-diffusion backbone, representing language, observations, goals, and actions as discrete tokens. One model learns action prediction, action-conditioned next-observation prediction, terminal goal-state prediction, and trajectory-to-instruction reconstruction, enabling test-time scaling through goal prediction and joint refinement. It is continually pretrained on approximately 1.33 million trajectories from 48 Open X-Embodiment datasets and achieves competitive performance on LIBERO and zero-shot LIBERO-Plus plus a 78.4% average success rate across four manipulation conditions on a Franka Research 3 robot. An optimized block-parallel implementation accelerates model-side action decoding by up to 29.2x.

  • Single diffusion model learns action, dynamics, goal, and instruction reconstruction objectives.
  • Pretrained on ~1.33M trajectories from 48 Open X-Embodiment datasets.
  • 78.4% average success across four conditions on Franka Research 3.
  • Competitive on LIBERO and zero-shot LIBERO-Plus.
  • Block-parallel decoding gives up to 29.2x speedup.
Full article239 words · extracted from arxiv.org · click to collapse

Visual goal and dynamics prediction can provide language-conditioned robot policies with both a target outcome and a representation of action-dependent scene changes. We bring these predictions into action generation and selection through a shared trajectory model. Dynin-Robotics implements this formulation on Dynin-Omni, an omnimodal masked-diffusion backbone, representing language, visual observations, goals, and actions as discrete tokens. By varying conditioning and target spans, the same model learns action prediction, action-conditioned next-observation prediction, terminal goal-state prediction, and trajectory-to-instruction reconstruction. These interfaces support test-time scaling through goal prediction, action-candidate evaluation, and joint refinement of action and future-state predictions. We continually pretrain the model on approximately 1.33 million trajectories from 48 Open X-Embodiment datasets and adapt it separately to downstream domains. On two VLABench tasks, robot pretraining improves adaptation within a fixed Stage-2 step budget, and the full objective mixture improves shifted-instruction success over Policy-only post-training under the same coupled decoder. Combining goal guidance with joint action-next-state denoising further improves shifted-instruction success over action-only decoding; the benefit depends on how the predictions are composed. Dynin-Robotics achieves competitive performance on LIBERO and zero-shot LIBERO-Plus, together with a 78.4% average success rate across four manipulation conditions on a Franka Research 3 robot. An optimized block-parallel implementation accelerates model-side action decoding by up to 29.2x relative to the base implementation under the reported profiling setup. These results support shared trajectory modeling as a common interface for learning complementary robot objectives and composing their predictions during control.

Text extracted automatically; images, tables and formatting may be missing. Original: https://arxiv.org/abs/2609.13053