KelinLiIC/Immersive-demonstrations

Immersive Demonstrations (ICRA 2023)

User-study demonstrations for the paper Immersive Demonstrations are the Key to Imitation Learning (Kelin Li, Digby Chappell and Nicolas Rojas, ICRA 2023). Code: github.com/KelinLi-ReCAST/Immersive-demonstrations.

Ten participants picked up a duck and placed it in a tray in PyBullet, driving a simulated end-effector with a SenseGlove and an HTC Vive Tracker. Each participant made 5 successful demonstrations per end-effector and condition.

End-effector	Folder
20-DoF human hand	mano
Franka parallel gripper	franka
RUTH hand	RUTH

Condition	Folder
Force feedback to the fingers	force
No force feedback	no_force
Force feedback to the fingers and, through a Franka Panda, to the wrist	panda

Download

The layout matches the code repository, so the files can be downloaded straight into a clone:

git clone https://github.com/KelinLi-ReCAST/Immersive-demonstrations.git
cd Immersive-demonstrations
hf download KelinLiIC/Immersive-demonstrations --repo-type dataset --local-dir . --include "data/*" --include "media/videos/*"

Files

Path	Content
data/<gripper>/<condition>/<participant>_<k>.npy	One row per control cycle: wrist force (3), wrist position (3), wrist Euler orientation (3), end-effector joint targets (mano: 20, RUTH: 3, franka: 1)
data/<gripper>/<condition>/finger_force_<participant>_<k>.npy	Force rendered to each of the 5 fingers, per cycle
data/<gripper>/glove_calibration/<participant>.npy	Glove calibration: range (21 values) and minimum (21 values) of each channel
data/<gripper>/training_results/	Forces and success rates read by evaluation.py for the plots of the paper
data/<gripper>/test_{force,noforce}_<gripper>.npy	Forces recorded while testing the behaviour-cloning policies (behavioral_cloning/test_bc_grasp.py), read by evaluation_wrist.py
data/<gripper>/ax.txt	Text dump of the last glove calibration
media/videos/<gripper>/<condition>/<participant>_<k>.mp4	PyBullet screen recording of the trial
data/Participants.txt	Participant numbers 1–10 used in the file names (anonymised)

The number of recordings per folder differs slightly from the number of demonstrations.

Application tasks (upgraded platform)

media/real_robot/ holds experiments on an upgraded version of the platform, in which the demonstrator, wearing a VR headset and a haptic glove, teleoperates a real robot arm with a dexterous hand. Each task folder has the robot recording, the demonstrator recording and both clips aligned in time, with the demonstrator as an inset (real time).

Folder	Task	Files
media/real_robot/pick_tomato/	Pick Tomato	robot.mp4 (1920 × 1080, 15.3 s) and operator.mp4 (15.4 s), both starting at the same moment; aligned.mp4
media/real_robot/battery_insertion/	Battery Insertion	robot.mp4 (1920 × 1080, 21.3 s) and operator.mp4 (21.1 s), both starting at the same moment; aligned.mp4
media/real_robot/plug_msd/	Plug MSD	robot.mp4 (1920 × 1080, 14.8 s); operator.mp4 (28.2 s, its 8.6 s mark corresponds to the start of robot.mp4); aligned.mp4
media/real_robot/press_button/	Press Button	robot.mp4 (1920 × 1080, 15.7 s) and operator.mp4 (15.7 s), both starting at the same moment; aligned.mp4

hf download KelinLiIC/Immersive-demonstrations --repo-type dataset --local-dir . --include "media/real_robot/*"

Application task demonstrations

application_tasks/ holds the teleoperated demonstrations of the four application tasks, recorded with the upgraded platform, with and without force feedback to the demonstrator. Each P0–P4 folder is a LeRobot dataset (format v2.1, 30 Hz):

application_tasks/<task>/<force_feedback|no_force_feedback>/P<k>/
  meta/        info.json, tasks.jsonl, episodes.jsonl, episodes_stats.jsonl, modality.json
  data/chunk-000/episode_<i>.parquet

Task	Folder	Episodes with force feedback	Episodes without force feedback
Pick Tomato	pick_tomato	53	55
Battery Insertion	battery_insertion	31	48
Plug MSD	plug_msd	45	62
Press Button	press_button	39	49

Each frame contains:

Column	Shape	Content
observation.state	8	Joint positions: 7 arm joints and the end-effector
observation.qvel	8	Joint velocities
observation.effort	8	Joint efforts
observation.force_torque	6	Wrist force/torque sensor (NaN where the sensor was not recorded)
action	8	Commanded joint positions
timestamp, frame_index, episode_index, index, task_index	1	Indices, as in LeRobot

The camera streams are not included.

hf download KelinLiIC/Immersive-demonstrations --repo-type dataset --local-dir . --include "application_tasks/*"

Citation

@inproceedings{li2023immersive,
  title     = {Immersive Demonstrations are the Key to Imitation Learning},
  author    = {Li, Kelin and Chappell, Digby and Rojas, Nicolas},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year      = {2023},
  note      = {arXiv:2301.09157}
}