⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

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

arXiv 2026 36.9 method

TLDR

Dynin-Robotics introduces a unified diffusion VLA model jointly predicting actions, future observations, goals, and instructions, improving robot policies via goal-guided action generation and selection.

Reasoning

The paper presents a novel unified formulation that couples action and dynamics prediction, with strong evaluation across multiple benchmarks and datasets. However, the abstract is truncated and provides limited details on real-world deployment, baseline comparisons, and the precise conditions under which composed prediction helps.

Read-first score

Read-first score 36.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 34.

Recency 5%
100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%
80

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,evaluation,result

Topical relevance 25%
48.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%
46

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code,dataset

Graph prestige 15%
5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%
0

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%
0

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 580.

Keyword Scores

world dynamics prediction
8
world model
6
interactive world model
6
generative world model
5
world simulator
4
video world model
3
model-based reinforcement learning world model
2

Deep Analysis

Innovations

  • Shared trajectory model unifying visual goal prediction, dynamics prediction, and action generation/selection for language-conditioned robot policies.
  • Dynin-Omni omnimodal masked-diffusion backbone representing language, visual observations, goals, and actions as discrete tokens.
  • Single model trained via varying conditioning and target spans to handle action prediction, action-conditioned next-observation prediction, terminal goal-state prediction, and trajectory-to-instruction reconstruction.
  • Test-time scaling interfaces including goal prediction, action-candidate evaluation, and joint refinement of action and future-state predictions.
  • Continual pretraining on approximately 1.33 million trajectories from 48 Open X-Embodiment datasets with separate downstream adaptation.
  • Block-parallel implementation accelerating model-side action decoding by up to 29.2x.

Methodology

Dynin-Robotics uses the Dynin-Omni omnimodal masked-diffusion backbone, tokenizing language, visual observations, goals, and actions as discrete tokens. The same model is trained by varying conditioning and target spans across action prediction, action-conditioned next-observation prediction, terminal goal-state prediction, and trajectory-to-instruction reconstruction. It is continually pretrained on approximately 1.33 million trajectories from 48 Open X-Embodiment datasets, adapted separately to downstream domains, and evaluated on VLABench, LIBERO, LIBERO-Plus, and a Franka Research 3 robot.

Key Results

Robot pretraining improves adaptation within a fixed Stage-2 step budget on two VLABench tasks, and the full objective mixture improves shifted-instruction success over Policy-only post-training under the same coupled decoder. Goal guidance combined with joint action-next-state denoising further improves shifted-instruction success over action-only decoding, with competitive LIBERO and zero-shot LIBERO-Plus performance, 78.4% average success across four Franka Research 3 manipulation conditions, and up to 29.2x faster action decoding from the optimized block-parallel implementation.

Limitations

  • The abstract notes that the benefit of combining goal guidance with joint action-next-state denoising depends on how the predictions are composed.

Resources

Paper

Tags