MetaWAM-RoboCasa-Predict2-2B

Paper | Code | From Video Frames to Robot Meta-Frames: A Unified Latent Interface for World Action Models (CoRL 2026)

MetaWAM is a World Action Model fine-tuned from Cosmos-Predict2-2B-Video2World. Robot signals (task type, proprioception, action chunks, future proprioception and values) are packed into fixed coefficient slots of disjoint channel bands and lifted into compact meta-frames with a 2D inverse DCT. Slot-aware masks then train policy, world-model and value prediction in a single denoising model.

This checkpoint is trained on RoboCasa (24 kitchen tasks, 50 demonstrations per task).

Files

File	Description
MetaWAM-RoboCasa-Predict2-2B.pt	Model weights (bf16, consolidated)
robocasa_dataset_statistics.json	Action / proprio normalization statistics
robocasa_t5_embeddings.pkl	Precomputed T5 embeddings of the task instructions
SHA256SUMS	Checksums

The statistics and T5 embedding files are identical to the ones released with Cosmos Policy.

Usage

Use the inference config metawam_robocasa__inference from the MetaWAM repository:

hf download Ethanliang99/MetaWAM-RoboCasa-Predict2-2B --local-dir checkpoints/MetaWAM-RoboCasa-Predict2-2B
bash scripts/test_robocasa.sh   # 24 tasks x 50 trials, seeds 195-197

See ROBOCASA.md for details.

License

These weights are fine-tuned from nvidia/Cosmos-Predict2-2B-Video2World; use is subject to the license of the base model.

Citation

@article{liang2026metawam,
  title   = {From Video Frames to Robot Meta-Frames: A Unified Latent Interface for World Action Models},
  author  = {Liang, Guoqiang and Lin, Yiqi and Chen, Yanzhe and Ma, Kevin Yuchen and Shou, Mike Zheng},
  journal = {arXiv preprint arXiv:2611.14365},
  year    = {2026}
}