Efficient-Large-Model/Long-WAM-RoboCasa-GR1-0s

Long-WAM RoboCasa-GR1 0s

Long-WAM policy weights for RoboCasa GR1 with 0 seconds of observation history. The duration denotes the history window, not the model parameter count or action chunk.

Download

hf download Efficient-Large-Model/Long-WAM-RoboCasa-GR1-0s --local-dir ./weights/Long-WAM-RoboCasa-GR1-0s

History and inference

P0 uses only the current observation, with no past history (one clean latent frame).

Load model.pt with the matching Long-WAM RoboCasa GR1 runtime and config.yaml. Use the supplied dataset_stats.json for the GR1 action normalization and state/action adapter. Input is the single ego_view RGB camera (256×256, preprocessed to 224×224), with 58-D model proprioception and 29-D actions.

Generate two future-video latents using 4 video denoising steps (sigma 0.9), then predict a 16-step action chunk using 10 action denoising steps. Execute 16 actions before replanning. Record observations at every 20-Hz control step; repeat the earliest available frame when history is short, and reset history at each episode boundary.

The matching runtime, GR1 simulator/assets, Wan VAE and task text embeddings are required separately. This release contains weights and inference settings only, without optimizer states or training artifacts.