MUSIC: Learning Muscle-Driven Dexterous Hand Control

This repository holds the data for MUSIC: Learning Muscle-Driven Dexterous Hand Control: the reference motions for hand-motion tracking, and the pretrained policies.

[Code] [Webpage] [arXiv] [YouTube] [TOG]

The code is at github.com/xupei0610/music. It reads the files from this repository without any conversion.

Contents

left_hand_motions/<id>.joblib        148 files, 237 clips, 651.6 min
right_hand_motions/<id>.joblib       153 files, 284 clips, 651.3 min
pretrained/
  joint_fingering/<piece>_fingering.ckpt     15 joint-driven piano policies, given fingering
  joint_hand/<piece>_hand.ckpt               15 joint-driven piano policies, hand only
  muscle_tracking/muscle_tracking_left.ckpt  left-hand muscle-driven tracking policy
  muscle_tracking/muscle_tracking_right.ckpt right-hand muscle-driven tracking policy

TODO: Pretrained high-level muscle-driven piano policies (cfg/muscle.py) are not included yet. We will add them in a later update.

Download

Use the script in the code repository. It writes the motions to motions/ in the repository root:

scripts/download_motions.sh                    # all motions
scripts/download_motions.sh --hand right       # one hand
scripts/download_motions.sh --files 000 001    # selected files, both hands

Or download with the Hugging Face CLI:

# all motions, about 8.7 GB
uvx --from huggingface-hub hf download xupei0610/MUSIC --repo-type dataset \
  --include "left_hand_motions/*" "right_hand_motions/*" --local-dir motions

# all pretrained policies, about 1.0 GB, into the pretrained/ folder of the code repository
uvx --from huggingface-hub hf download xupei0610/MUSIC --repo-type dataset \
  --include "pretrained/*" --local-dir .

Reference motions

The motions are hand motions from piano performances, retargeted from the ForElise dataset to the MUSIC hand model. Each hand has about 651 minutes of motion at 60 frames per second. The tracking task (cfg/muscle_tracking_{left,right}.py) and the piano tasks use them.

File format

Each .joblib file holds a list of clips. Each clip is a Motion named tuple from ref_motion.py:

Motion = namedtuple("Motion", "fps pos orient ang_vel lin_vel joint_pos joint_vel")

Field	Shape	Meaning
fps	int	Frame rate, always 60.
pos	(T, 34, 3)	Position of each body in the world frame, in meters.
orient	(T, 34, 4)	Orientation of each body in the world frame, as a unit quaternion (x, y, z, w).
ang_vel	(T, 34, 3)	Angular velocity of each body in the world frame, in rad/s.
lin_vel	(T, 34, 3)	Linear velocity of each body in the world frame, in m/s.
joint_pos	(T, 23)	Angle of each hinge joint of the hand, in radians.
joint_vel	(T, 23)	Angular velocity of each hinge joint, in rad/s.

T is the number of frames of the clip, and all arrays are torch.float32 tensors. The 34 bodies are the bodies of assets/{left,right}_hand_muscle_driven.xml in model order, without the world body. The first bodies are ulna, radius and lunate. The 23 joints are the hinge joints of the same model in model order, without the free root joint.

Two files, right_hand_motions/050.joblib and right_hand_motions/111.joblib, contain an empty list. The loader skips them.

Load a file

The files store the Motion type of ref_motion.py, so ref_motion must be importable. Run the following from the root of the code repository:

import joblib
from ref_motion import Motion   # the type that the files store

clips = joblib.load("motions/left_hand_motions/000.joblib")
clip = clips[0]
print(clip.fps, clip.pos.shape, clip.joint_pos.shape)

Pretrained policies

All policy files keep only what running the policy needs: the actor weights and the observation normalizer mean and variance. They contain no critic, discriminator, optimizer state or training records. Each file is a dict of tensors, so torch.load(path, weights_only=True) can read it.

Joint-driven piano policies

Each file controls both hands for one piece, with joint position targets. It plays the note file with the same name from the notes/ folder of the code repository. For example, pretrained/joint_hand/052-1_hand.ckpt plays notes/052-1_hand.txt.

• joint_fingering/ was trained with the fingering of the PIG dataset, one finger for each note.
• joint_hand/ was trained with only the hand of each note. The policy selects the finger.

The policies cover 15 pieces:

PIG ID	Piece
017-1	Mozart, Piano Sonata K. 545 in C major, 1st movement
046-1	Bach, French Suite No. 3, Minuet
048-1	Bach, English Suite No. 2, Prelude
052-1	Mozart, Piano Sonata K. 280 in F major, 2nd movement
061-2	Beethoven, Für Elise
070-1	Beethoven, Piano Sonata No. 23, 2nd movement
075-1	Chopin, Waltz Op. 64 No. 2
087-1	Debussy, Suite bergamasque, Passepied
092-1	Grieg, Lyric Pieces Op. 38 No. 1
102-1	Mussorgsky, Pictures at an Exhibition, Bydło
117-1	Brahms, Rhapsody Op. 79 No. 2
119-1	Joplin, Maple Leaf Rag
123-1	Beethoven, Piano Sonata No. 3, 2nd movement
140-1	Scarlatti, Sonata in A major K. 208
145-1	Scriabin, Piano Sonata No. 5

Mean scores over the 15 pieces, from one full play of each piece with the mean action:

Policies	Frame F1	Note F1
joint_fingering	0.990	0.983
joint_hand	0.986	0.982

Frame F1 compares the pressed keys with the score at each 60 Hz control frame. Note F1 matches a played note to a score note on the same key when the two onsets are within 50 ms of each other.

Evaluate a policy in the viewer:

uv run python main.py cfg/joint.py --test \
  --note notes/017-1_fingering.txt \
  --ckpt pretrained/joint_fingering/017-1_fingering.ckpt

Muscle-driven tracking policies

muscle_tracking_left.ckpt and muscle_tracking_right.ckpt are the general low-level tracking policies for each hand. They control 44 muscles and the 6 root force and torque actuators of the hand at 480 Hz, and they track the reference motions above. The high-level muscle-driven piano task (cfg/muscle.py) uses them as its controllers.

TODO: Pretrained high-level muscle-driven piano policies (cfg/muscle.py) are not included yet. We will add them in a later update.

Evaluate a tracking policy:

uv run python main.py cfg/muscle_tracking_right.py --test \
  --ckpt pretrained/muscle_tracking/muscle_tracking_right.ckpt

License

This repository is released under the MIT License. The source datasets keep their own licenses. See the data attribution below.

Data attribution

The reference motions are retargeted from the ForElise piano performance dataset. The note files are extracted from the Piano Fingering Dataset (PIG). Please follow the licenses and citation requirements of the source datasets.

Citation

If you use this data, please cite:

@article{music,
  author = {Xu, Pei and Ye, Yufei and Sun, Shuchun and Ding, Yu and Schumann, Elizabeth and Liu, C. Karen},
  title = {{MUSIC}: Learning Muscle-Driven Dexterous Hand Control},
  journal = {ACM Transactions on Graphics},
  volume = {45},
  number = {4},
  year = {2026},
  doi = {10.1145/3811402}
}