DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection

Official implementation of DexPort.

[Paper] [Website]

Installation

python3.11 -m venv .venv && source .venv/bin/activate
pip install -e ".[collect]"   # or train, deploy, calib, wandb

Do the per-machine setup in docs/usage.md first: git-lfs before pip for the Quest reader, CUDA torch for training and deployment, and the Marvin SDK for the Tianji arms.

Hardware

hardware/ has the 3D-printed parts of the hand-held interface (STL, printed in PLA) and the bill of materials, BOM.csv:

part	files
handle	handle_left.stl, handle_right.stl
camera mount, Wuji 1	cam_mount_left.stl, cam_mount_right.stl
camera mount, Wuji 2	cam_mount_left_wuji2.stl, cam_mount_right_wuji2.stl
strap bracket	strap_bracket.stl

Training the Rotation Skill in Simulation

We train the in-hand rotation skill with PPO in Isaac Lab, which needs its own conda environment:

bash sim/install.sh            # conda env "isaaclab"
conda activate isaaclab
python sim/train.py --headless

checkpoints/rotation_skill/model.pt is a random placeholder: replace it with a trained actor before collecting screwdriver demonstrations. See sim/README.md.

Data Collection

Run the collector with one of the task configs in configs/collect/ (wuji1_pnp, wuji1_screwdriver, wuji2_drawer, wuji2_scan):

python -m dexport.collect --config configs/collect/wuji1_pnp.yaml

On the Quest controllers, B starts and stops recording and Y discards the episode; each run saves a session to data/<MMDD_HHMM>_<task>/. The full controls are in docs/usage.md.

Policy Training

Split the sessions into train and test, then train a diffusion policy with one of the presets (pnp, pnp_cotrain, screwdriver, drawer, scan):

python scripts/make_split.py --sessions data/*_pnp --out data/splits/pnp --test-count 10
python -m dexport.minbc.train pnp --data-dir data/splits \
    --train-data pnp/train/trajectories.zarr --test-data pnp/test/trajectories.zarr

python -m dexport.minbc.train <preset> --help lists a preset's fields; any of them can be overridden on the command line.

Real-World Deployment

Calibrate the controller-to-flange mount of the policy's arm (A is left, B is right), paste the fit into dexport/deploy/config.py, then run the policy on the Tianji arms:

python -m dexport.calib.capture --mode jog --arm B --out data/calib/arm_b.npz
python -m dexport.calib.solve --npz data/calib/arm_b.npz   # prints the QUEST_TO_* lines to paste
python -m dexport.deploy --ckpt outputs/pnp

Before the first run on a rig, see docs/usage.md for the calibration procedure, --spacer and replaying a recording.

Citation

@misc{xie2026dexport,
  title  = {DexPort: A Hand-Held Skill Interface for Dexterous Hand Data Collection},
  author = {Amber Xie and Jiawei Fu and Xiaofeng Guo and Carlo Sferrazza and Guanya Shi and Haozhi Qi},
  year   = {2026},
  url    = {https://amberxie88.github.io/dexport/index.html}
}

Security

See CONTRIBUTING for more information.

License

This code is released under the Apache License 2.0. Third-party notices are in THIRD_PARTY_NOTICES.md.