DexWM: World Models for Learning Dexterous Hand-Object Interactions from Human Videos
_{Official PyTorch Implementation}

Paper | Project Page | Data (RoboCasa Random)

This repo contains the official PyTorch implementation of DexWM: World Models for Learning Dexterous Hand-Object Interactions from Human Videos.

Authors:
Raktim Gautam Goswami^1,2, Amir Bar¹, David Fan¹, Tsung-Yen Yang¹, Gaoyue Zhou^1,2, Prashanth Krishnamurthy², Michael Rabbat¹, Farshad Khorrami², Yann LeCun^1,2

¹ Meta-FAIR ² New York University

Setup

Download the repo and set up the environment:

git clone https://github.com/facebookresearch/dexwm
conda create -n dexwm python=3.11
conda activate dexwm
pip install -r requirements.txt

Data

DexWM is pre-trained in EgoDex and DROID datasets and fine-tuned on exploratory sequences of the RoboCasa simulation data. Download the EgoDex, DROID, RoboCasa Random datasets. See the end of this README for the expected directory structure inside each dataset folder.

Training

Pre-Train on EgoDex and DROID

Note: Change the egodex_root_folder and droid_root_folder locations in the config file before running the code.

Using torchrun:

bash scripts/train_torchrun.sh --job_dir <job_dir>

Update the script variables to match your available compute resources (e.g., number of nodes, GPUs per node, and host address). Defaults are 1 node, 8 GPUs per node, and localhost.

Or using submitit and slurm:

bash scripts/train_submitit.sh

Update the script variables to match your available compute resources and job_dir. By default, this script trains the model on 32 nodes with 8 GPUs each.

Or locally on one GPU for debug:

python train_wm.py --config configs/egodex_and_droid.yaml --job_dir <job_dir>

On the first training run, the code generates split_indices_droid.json to define a DROID validation split. This file is only used to report/track validation loss and is not used elsewhere.

Fine-Tune on RoboCasa Random Data

Change the root_folder location and resume path to the pre-trained model in the config file before running the code.

Using torchrun:

bash scripts/multistep_train_torchrun.sh --job_dir <job_dir>

Update the script variables to match your available compute resources (e.g., number of nodes, GPUs per node, and host address). Defaults are 1 node, 1 GPUs per node, and localhost.

Or using submitit and slurm:

bash scripts/multistep_train_submitit.sh

Update the script variables to match your available compute resources and job_dir. By default, this script trains the model on 1 nodes with 1 GPUs each.

Or locally on one GPU for debug:

python train_multistep_wm.py --config configs/robocasa_random_multistep.yaml --job_dir <job_dir>

On the first training run, the code generates split_indices_robocasa_random.json to define a RoboCasa Random validation split. This file is only used to report/track validation loss and is not used elsewhere.

Evaluation

Rollout L2 Error and PCK on EgoDex

1. Set the model checkpoint: Edit test_scripts/test_script.sh and update the model checkpoint path to the checkpoint you want to evaluate.
2. Download the keypoint model: Evaluation also uses a separately trained keypoint model to predict keypoints from the world model’s predicted latent states. Download this model from the checkpoint download page and configure its path in test_scripts/test_script.sh as well.
3. (Optional) Visualization: The test script can visualize predicted states. To enable this, you must train a decoder and configure the decoder path/settings in the code.

Run evaluation

bash test_scripts/test_script.sh

This writes two rollout metrics to the output_dir specified in test_scripts/test_script.sh:

• L2 Error
• PCK (Percentage of Correct Keypoints)

Each metric is saved as an array evaluated every 0.2 seconds, from 0.2s up to 4.0s.

Compute summary statistics

To view the aggregated losses similar to the format reported in the paper, run

python test_scripts/result_stats.py --output_dir <output_dir>

Robot Manipulation Tasks

1. Install and configure RoboCasa simulator with MURP robot following the instructions here.
2. Download the Pick-and-Place dataset. It provides the visual goal images used for the manipulation tasks.
3. Run evaluation

   conda activate robot_sim_dexwm
   bash scripts/test_robot_sim.sh

   Before running, update the script variables to match your compute setup (e.g., number of nodes/GPUs), job_dir, and any other relevant settings. By default, the script uses 1 node with 8 GPUs.
4. At the end of evaluation, a res.json file will be generated in the job_dir which will contain a dictionary with all the task names and corresponding success/failure.

Dataset Directory Structure

The EgoDex and DROID datasets are aranged as follows:

egodex
├── train
│   ├── <task_1>
│   │   ├── 0.hdf5
│   │   ├── 0.mp4
│   │   ├── 1.hdf5
│   │   └── 1.mp4
│   │   ...
│   ├── <task_2>
│   │   ├── 0.hdf5
│   │   ├── 0.mp4
│   │   ├── 1.hdf5
│   │   └── 1.mp4
│   │   ...
│   ...
├── test
│   ├── <task_k>
│   │   ├── 0.hdf5
│   │   ├── 0.mp4
│   │   ├── 1.hdf5
│   │   └── 1.mp4
│   │   ...
│   ...

DROID
├── <lab_name>
│   ├── success
│   │   ├── <date_1>
│   │   │   ├── <time_1>
│   │   │   │   ├── recordings
│   │   │   │   │   ├── MP4
│   │   │   │   │   │   └── ...
│   │   │   │   │   ├── SVO
│   │   │   │   │   │   └── ...
│   │   │   │   ├── metadata_....json
│   │   │   │   └── ...
│   │   └── ...
│   │   ...
│   ├── failure
│   │   ├── <date_i>
│   │   │   ├── <time_j>
│   │   │   │   ├── recordings
│   │   │   │   │   ├── MP4
│   │   │   │   │   │   └── ...
│   │   │   │   │   ├── SVO
│   │   │   │   │   │   └── ...
│   │   │   │   ├── metadata_....json
│   │   │   │   └── ...
│   │   └── ...
│   │   ...

robocasa_random_data
├── exploratory_movements
│   ├── combine_demos_0.hdf5
│   └── combine_demos_1.hdf5
│   │   ...
├── gripper_open_and_close
│   ├── combine_demos_0.hdf5
│   └── combine_demos_1.hdf5
│   │   ...
├── pick-and-place-2.0
│   ├── combine_demos_0.hdf5
│   └── combine_demos_1.hdf5
│   │   ...

License

DexWM is licensed under CC-BY-NC.

BibTeX

@article{goswami2025world,
  title={World Models for Learning Dexterous Hand-Object Interactions from Human Videos},
  author={Goswami, Raktim Gautam and Bar, Amir and Fan, David and Yang, Tsung-Yen and Zhou, Gaoyue and Krishnamurthy, Prashanth and Rabbat, Michael and Khorrami, Farshad and LeCun, Yann},
  journal={arXiv preprint arXiv:2512.13644},
  year={2026}
}