Haptic-ACT with wrist force/torque: trained models

Models of the ManiSkill experiments of Haptic-ACT (section 8 of its README): Haptic-ACT with the 6-axis wrist force/torque instead of fingertip forces, and the baselines ACT, Diffusion Policy and BC, trained on the datasets of four tasks with a Franka Panda in ManiSkill. Each folder ft_maniskill/<task>/<method>/ is a LeRobot 0.6.1 pretrained_model directory (config, weights, pre/post-processors, training config).

Method	Folder	Policy type
Haptic-ACT (F/T)	ft	acthaptic (plugin)
ACT	act	act
Diffusion Policy	dp	diffusion
BC (ACT without chunking or CVAE)	bc	act

Each model is the checkpoint with the highest success over 10 seeds x 25 trials (for PushCube the last one). Success (%) of that checkpoint, mean ± std over the seeds:

Task	Haptic-ACT (F/T)	ACT	Diffusion Policy	BC
PegInsertionSide (peg)	79.6 ± 6.1 (180k)	43.6 ± 10.3 (190k)	78.4 ± 9.5 (150k)	1.2 ± 1.8 (70k)
PickCube (pickcube)	80.4 ± 6.3 (150k)	78.0 ± 6.5 (135k)	84.8 ± 6.1 (142.5k)	22.0 ± 8.8 (142.5k)
StackCube (stackcube)	87.2 ± 10.1 (127.5k)	82.0 ± 10.2 (127.5k)	95.6 ± 3.8 (142.5k)	0.0 ± 0.0 (150k)
PushCube (pushcube)	100.0 ± 0.0 (60k)	100.0 ± 0.0 (60k)	99.2 ± 1.6 (60k)	9.6 ± 6.0 (60k)

These numbers are of the selected checkpoints, chosen on the evaluation seeds, so they are optimistic. The results reported in the GitHub README average the last 10 checkpoints of each run.

Evaluation (installation in the GitHub README, section 8.3):

hf download KelinLiIC/Haptic-ACT --local-dir models_ft --include "ft_maniskill/*"
python ft_maniskill/eval/eval_ft.py --task peg --policy-path models_ft/ft_maniskill/peg/ft \
    --n-action-steps 50 --learned-trigger 0.5 --out outputs/eval/peg/ft

Citation

@inproceedings{li2025hapticact,
  title     = {Haptic-ACT: Bridging Human Intuition with Compliant Robotic Manipulation via Immersive VR},
  author    = {Li, Kelin and Wagh, Shubham M and Sharma, Nitish and Bhadani, Saksham and Chen, Wei and Liu, Chang and Kormushev, Petar},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2025},
  doi       = {10.1109/IROS60139.2025.11247471},
  note      = {arXiv:2409.11925}
}