--- library_name: pytorch license: other pipeline_tag: robotics tags: - long-wam - libero - idm --- # Long-WAM LIBERO-IDM Long-WAM policy weights for LIBERO. IDM predicts actions after generating future-video latents. For more details, refer to: - **Paper**: [Long-WAM: Scaling the Context of World-Action Models](https://huggingface.co/papers/2610.10528) - **Project page**: [Long-WAM project](https://nvlabs.github.io/LongLive/Long-WAM/) - **Code**: [GitHub repository](https://github.com/NVlabs/LongLive) ## Download ```bash hf download Efficient-Large-Model/Long-WAM-LIBERO-IDM --local-dir ./weights/Long-WAM-LIBERO-IDM ``` ## 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, 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. Use the matched LIBERO execution cadence. 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.