| library_name: pytorch | |
| license: other | |
| pipeline_tag: robotics | |
| tags: | |
| - long-wam | |
| - robocasa365 | |
| - idm | |
| # Long-WAM RoboCasa365 | |
| Long-WAM policy weights for RoboCasa365. | |
| IDM predicts actions after generating future-video latents. | |
| **Paper:** [Long-WAM: Scaling the Context of World-Action Models](https://huggingface.co/papers/2610.10528) | |
| **Project page:** [nvlabs.github.io/LongLive/Long-WAM](https://nvlabs.github.io/LongLive/Long-WAM/) | |
| **Code:** [github.com/NVlabs/LongLive](https://github.com/NVlabs/LongLive) | |
| ## Download | |
| ```bash | |
| hf download Efficient-Large-Model/Long-WAM-RoboCasa365 --local-dir ./weights/Long-WAM-RoboCasa365 | |
| ``` | |
| ## Inference | |
| Load `model.pt` with the matching Long-WAM runtime, using `config.yaml` and | |
| the supplied `dataset_stats.json` for preprocessing and normalization. | |
| P48 context uses 48 past control steps plus the current observation: **2.4 seconds at 20 Hz**, encoded into four clean latent frames. | |
| Generate two future-video latents using 4 steps (imagination sigma 0.9), then predict a 32-step action chunk using 10 action steps. | |
| Record observations at every control step and replan after executing 32 actions. Reset history at the start of each episode. | |
| The matching inference runtime and base-model VAE/text assets are required | |
| separately. This release contains weights and inference settings only; | |
| optimizer states and training artifacts are not included. |