Enfold Checkpoints

This repository provides the official checkpoints for Enfold: Folding World-Generator Computation into Predictive Representations for Efficient Embodied Control

What is included

Each benchmark has one policy checkpoint and one matching normalization-statistics file. Always use a checkpoint with the statistics file from the same row.

Benchmark	Policy checkpoint	Normalization statistics	Evaluation setup
LIBERO	enfold_libero.pt	libero_dataset_stats.pt	Two camera views, concatenated to 224 × 448; 7-D actions and 8-D proprioception
RoboTwin	enfold_robotwin.pt	robotwin_dataset_stats.pt	Three camera views at 384 × 320; 14-D actions and proprioception

Download

Install the Hub client, then download only the files for the benchmark you plan to evaluate:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="richardxyt/Enfold",
    local_dir="checkpoints/enfold_libero",
    allow_patterns=["enfold_libero.pt", "libero_dataset_stats.pt"],
)

For RoboTwin, replace the two filenames with enfold_robotwin.pt and robotwin_dataset_stats.pt.

Evaluation

First install the Enfold source code and its dependencies, then configure the required external model assets (Cosmos-Predict2.5, Cosmos tokenizer/VAE and text-encoder assets, DINOv3 ViT-H/16+). The source configuration contains paths that must be changed for your machine.

From the Enfold project root, evaluate the matching checkpoint/statistics pair:

# LIBERO
bash eval.sh libero \
  checkpoints/enfold_libero/enfold_libero.pt \
  checkpoints/enfold_libero/libero_dataset_stats.pt

# RoboTwin
bash eval.sh robotwin \
  checkpoints/enfold_robotwin/enfold_robotwin.pt \
  checkpoints/enfold_robotwin/robotwin_dataset_stats.pt

See the Enfold source repository for environment setup, dataset preparation, optional Hydra overrides, multi-GPU evaluation.

Citation

If these checkpoints help your research, please cite the model repository:

@software{enfold2026,
  title  = {Enfold: Robotics Policy Checkpoints},
  author = {Enfold Authors},
  year   = {2026},
  url    = {https://huggingface.co/richardxyt/Enfold}
}