Haptic-ACT: datasets and models

Datasets and trained models of Haptic-ACT: the MuJoCo experiments and real-robot rollouts of the paper, and the version with wrist force/torque in ManiSkill.

haptic_act/, act_baseline/: Haptic-ACT and ACT in MuJoCo

The data and models of the simulation experiments of the paper (sections 1-7 of the GitHub README): an xArm7 with the Inspire hand picks up an object and places it in a basket. The paths match the GitHub repository, so the download can go straight into a clone.

Path	Content
haptic_act/data/pick_and_place/raw_data/episode_<i>.hdf5	50 demonstrations, 400 steps each: observations/qpos, observations/qvel, action (400 x 13: 7 arm and 6 hand joints), observations/contact_force (400 x 5 fingertip forces, N), observations/images/frontview (400 x 480 x 640 x 3)
haptic_act/ckpt/	Haptic-ACT: policy_best.ckpt, dataset_stats.pkl, result_policy_best.txt and the training curves
act_baseline/ckpt/	ACT, trained on the same demonstrations: the same files

git clone https://github.com/KelinLi-ReCAST/Haptic-ACT.git && cd Haptic-ACT
hf download KelinLiIC/Haptic-ACT --repo-type dataset --local-dir . --include "haptic_act/*" --include "act_baseline/*"

ACT reads the demonstrations from haptic_act/data/: set DATA_DIR in act_baseline/constants.py to that folder (section 4 of the GitHub README).

ft_maniskill/: Haptic-ACT with wrist force/torque in ManiSkill

Demonstrations of five tasks with a Franka Panda with wrist camera in ManiSkill (section 8 of the GitHub README), recorded in the environments of the haptic-act branch of KelinLi-ReCAST/ManiSkill. Each ft_maniskill/<task>/lerobot_ft/ is a LeRobot dataset (format v3.0, 20 fps), and ft_maniskill/<task>/models/ holds the models trained on it.

Task	Folder	ManiSkill environment	Demonstrations	Episodes	Frames
PegInsertionSide	peg	PegInsertionSideFT-v1	jam and recovery	1000	196,523
PickCube	pickcube	PickCubeVisGoal-v1	motion planner	1000	78,246
StackCube	stackcube	StackCube-v1	ManiSkill's demonstrations, replayed	1000	107,420
PushCube	pushcube	PushCube-v1	ManiSkill's demonstrations, replayed	990	68,287
PlugChargerTopDown	plug_topdown	PlugChargerTopDown-v1	jam and recovery	1000	173,314

Column	Shape	Content
observation.images.base_camera, observation.images.wrist_camera	480 x 480 x 3	RGB video
observation.state, observation.qvel	9	joint positions / velocities (7 arm joints, 2 fingers)
action	8	absolute targets of the 7 arm joints and the gripper (pd_joint_pos)
observation.torque	7	arm-joint torques (N m)
observation.wrench	6	TCP wrench [fx fy fz tx ty tz] (N, N m, TCP frame), least squares of tau = J^T F
observation.environment_state	30	the last 5 wrenches, oldest first: the F/T input of Haptic-ACT
observation.torque_hist	35	the last 5 joint-torque readings
observation.torque_do	7	joint torques with 30 % of the frames zeroed
observation.peg_box_force (peg) / observation.contact_force	1	contact force between the object and the box / goal (N), privileged, not a policy input
phase	1	demonstration phase (peg: reach, grasp, align, jam, backoff, realign, insert)
replan	1	1 during the deliberate jam of the peg demonstrations (contact force > 2 N), else 0
next.success	1	task success after the frame's action

The baselines (ACT, Diffusion Policy, BC) train on a copy without the force columns, made by ft_maniskill/dataset_processing/make_baseline_dataset.py of the GitHub repository.

Models

The models of PlugChargerTopDown will be added.

Each ft_maniskill/<task>/models/<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)	haptic-act	acthaptic (plugin)
ACT	act	act
Diffusion Policy	dp	diffusion
BC (ACT without chunking or CVAE)	bc	act

Download and evaluation

hf download KelinLiIC/Haptic-ACT --repo-type dataset --local-dir data_ft --include "ft_maniskill/*"
# only the datasets: --include "ft_maniskill/*/lerobot_ft/*"; only the models: --include "ft_maniskill/*/models/*"
python ft_maniskill/eval/eval_ft.py --task peg --policy-path data_ft/ft_maniskill/peg/models/haptic-act \
    --n-action-steps 50 --learned-trigger 0.5 --out outputs/eval/peg/haptic-act

Installation, training and evaluation of every method: section 8 of the GitHub README.

real_robot/: real-robot rollouts of Haptic-ACT and ACT

Rollouts of the trained policies on the real robot (section 6 of the GitHub README): Haptic-ACT in real_robot/real_hact/ and ACT in real_robot/real_act/, 50 episodes each. Each episode_<i>_compressed.hdf5 has 400 steps:

Key	Content
observations/qpos, observations/qvel, observations/effort	400 x 13: 7 arm joints and 6 hand joints
action	400 x 13 joint targets
observations/images/end_effector, observations/images/frontview	JPEG-encoded 480 x 640 frames, one per step, zero-padded to a fixed length
compress_len	2 x 400 encoded lengths; row 0 is end_effector, row 1 is frontview

real_robot/data_extract.py of the GitHub repository reads these files (load_hdf5).

hf download KelinLiIC/Haptic-ACT --repo-type dataset --local-dir . --include "real_robot/*"

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}
}