| --- |
| license: other |
| license_name: nvidia-open-model-license |
| base_model: nvidia/Cosmos-Predict2-2B-Video2World |
| library_name: cosmos |
| pipeline_tag: image-to-video |
| tags: |
| - robotics |
| - world-model |
| - action-conditioned-video |
| - cosmos |
| --- |
| |
| # Prophet World Model |
|
|
| Built on NVIDIA Cosmos. |
|
|
| Prophet is the single-view, history-aware, action-conditioned robot world model introduced in **Reinforcing Action Policies by Prophesying**. It predicts future robot-observation videos from an initial observation, a 20-step action chunk, and an optional history buffer and projected action-frame sequence. |
|
|
| ## Released Variants |
|
|
| - `prophet-pretrained.pt`: heterogeneous robot-data pretrained checkpoint used for downstream adaptation. |
| - `prophet-bridge.pt`: full-parameter BRIDGE-adapted checkpoint. |
| - `prophet-libero.pt`: full-parameter LIBERO-adapted checkpoint. |
|
|
| Exact sizes and SHA-256 checksums are listed in `weights.json`. |
|
|
| Internal run identifiers, training dates, and exact checkpoint iterations are intentionally omitted from the public release metadata. |
|
|
| ## Base Model and License |
|
|
| The checkpoints are derivative models of `nvidia/Cosmos-Predict2-2B-Video2World`. They are governed by the NVIDIA Open Model License included in `NVIDIA_OPEN_MODEL_LICENSE.md`. The repository source code is separately available under Apache-2.0. |
|
|
| Licensed by NVIDIA Corporation under the NVIDIA Open Model License. |
|
|
| ## Intended Use |
|
|
| - Research on action-conditioned robot world models. |
| - Open-loop or autoregressive video rollout generation. |
| - Initialization for BRIDGE, LIBERO, or similarly formatted robot-data adaptation. |
| - Learned-simulator research with independent real-environment validation. |
|
|
| ## Limitations |
|
|
| - Generated rollouts can hallucinate geometry, contacts, object state, or task success. |
| - Long autoregressive rollouts accumulate visual and dynamics errors. |
| - The model is not a replacement for real-robot safety validation or a certified simulator. |
| - Dataset-specific camera calibration and action semantics affect transfer quality. |
| - The released checkpoints are single-view; A2World multi-view models are not included. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @article{zhang2025prophrl, |
| title={Reinforcing Action Policies by Prophesying}, |
| author={Zhang, Jiahui and Huang, Ze and Gu, Chun and Ma, Zipei and Zhang, Li}, |
| year={2025}, |
| journal={arXiv preprint arXiv:2511.20633}, |
| } |
| ``` |
|
|
