| --- |
| license: mit |
| tags: |
| - robotics |
| - reinforcement-learning |
| - world-model |
| - libero |
| - video-generation |
| --- |
| |
| # Pretrained pixel world-model assets |
|
|
| Anonymous supplementary checkpoints for the accompanying double-blind |
| submission. These assets support pixel-based tree-search experiments on five |
| LIBERO-90 tasks. |
|
|
| ## Contents |
|
|
| - `checkpoint_00200000.safetensors`: action-conditioned, two-view, 128-pixel |
| video world model trained for 200,000 steps. |
| - `t5_context_cache/`: precomputed frozen text-encoder outputs for the five |
| released task prompts and the empty negative prompt. |
| - `onestep_state_wm/.../checkpoint_best.safetensors`: one-step state model |
| mapping `(state, action)` to the next state. |
|
|
| The base Wan2.2 VAE and tokenizer are third-party assets and are intentionally |
| not redistributed here. Follow the anonymous code artifact README to obtain |
| them from their official source. |
|
|
| Download all released assets with: |
|
|
| ```bash |
| hf download qlearningwithworldmodels/anonymous-pixel-world-models \ |
| --local-dir pixel_world_models |
| ``` |
|
|
| Verify downloads against `SHA256SUMS` before use. Author and archival links |
| are intentionally omitted during double-blind review and will be added after |
| the review period. |
|
|
