A2World-World-Model

A2World World Model

A2World is a multi-view action-conditioned diffusion world model trained to predict future robot observations from initial camera observations and a 20-step action chunk. A2World-sim adds pose-guided visual history and autoregressive rollout support.

Variants

• a2world-pretrained.pt: heterogeneous robot-data pretrained A2World checkpoint.
• a2world-libero.pt: history-aware A2World-sim checkpoint adapted on LIBERO.

Both released files use EMA weights, the standard net.* inference prefix, and BF16 storage compatible with the public rollout pipeline.

Base Model and License

The checkpoints are derivative models of nvidia/Cosmos-Predict2-2B-Video2World and are governed by the NVIDIA Open Model License included in this repository. Source code is separately available under Apache-2.0.

Intended Use

• Multi-view action-conditioned robot video prediction.
• Open-loop and chunk-wise autoregressive world-model rollouts.
• LIBERO learned-simulator research and downstream adaptation.
• Research on transferable robot dynamics priors.

Limitations

• Generated rollouts may hallucinate geometry, contact, object state, or task success.
• Autoregressive errors accumulate over long horizons.
• Camera identity and action semantics must match the selected embeddings and preprocessing.
• The model is not a certified simulator and must not replace real-robot safety validation.

Citation

@inproceedings{huang2026a2world,
    title={Learning Transferable Dynamics Priors from Action to World Modeling},
    author={Huang, Ze and Zhang, Jiahui and Liu, Hairuo and Zhang, Chenxi and Cheng, Ran and Zhang, Li},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    year={2026},
}