NowWAM

Weights for Beyond Future Prediction: Denoising as Generative Adaptation for Robot Control. Code · Project

All three checkpoints are JAX-trained EMA weights converted to PyTorch BF16. Each folder includes model.pth, dataset_stats.json, config.json, provenance.json, and a model card with the reference evaluation results.

Folder	Backbone	Policy	Benchmark
Klein-LIBERO	Klein, 15 layers	Regression, 1 step	LIBERO-Plus
ZImage-LIBERO	Z-Image, 30 layers	Regression, 1 step	LIBERO-Plus
Klein-RoboCasa-100shot	Klein, 25 layers	Flow, 4 steps	RoboCasa GR1, 100-shot

Download only the checkpoint you need:

from huggingface_hub import snapshot_download
import torch

folder = "Klein-LIBERO"  # or ZImage-LIBERO / Klein-RoboCasa-100shot
path = snapshot_download("xmz111/NowWAM", allow_patterns=[f"{folder}/*"])
state = torch.load(f"{path}/{folder}/model.pth", map_location="cpu",
                   weights_only=True, mmap=True)

The checkpoint contains mot and proprio_encoder state dictionaries. Text encoders and VAEs must be obtained separately from their upstream models. The standalone evaluation package is being prepared; a validated clean-install rollout recipe is not yet included. Reference scores and evaluation caveats are documented in the individual model cards. Consolidating these files does not change their evaluation settings or establish new results.

manifest.json records checkpoint checksums and source revisions. Weights are distributed under Apache-2.0. Upstream models, simulation assets and software retain their respective licenses. The Torch implementation builds on ImageWAM.