Lite Motion Retargeting

Retargets the LAFAN1 motion capture set from the Unitree G1 to the Berkeley Lite humanoid. One clip becomes one layered Rerun motion in Berkeley-Humanoids/Lite-LAFAN1-Dataset, which you can open in a browser.

Use

uv sync
uv run scripts/download_lafan.py         # G1 clips -> .cache/
uv run scripts/retarget.py --workers -1  # every clip, every core
uv run rerun dataset/*/lafan1_walk1_subject1.rrd dataset/blueprints/lite_pro.rbl
uv run scripts/publish.py --clear        # upload to the Hub

To write to the Hub, run hf auth login or set HF_TOKEN.

Script	What it does
download_lafan.py	Downloads the G1 clips of LAFAN1 into .cache/lafan1_g1/.
retarget.py	Retargets each clip and writes its two layers under dataset/.
store.py	Defines the file format.
publish.py	Fills the motion index of DATASET_CARD.md, then uploads both to the Hub.
common.py	The joint table, the model loaders, the CSV reader, and forward kinematics.

retarget.py takes --clip <regex>, --no-ik, --ik-iters N, --workers N, and --validate-only, which prints the end-effector error table and writes nothing. publish.py takes --dry-run, --repo-id <id>, and --clear, which deletes every file in the Hub repo that this store does not replace.

How a clip is retargeted

The source is 30 Hz and a motion is written at 50 Hz, the control rate of the training environment. retarget.py resamples the source before it does anything else, so the IK solves every frame that reaches the file, and the body poses stay the exact kinematics of the joints beside them.

1. Direct remap. lite_q = sign * g1_q + offset, from the table in common.G1_TO_LITE. A constant shift of the pelvis height puts the feet of Lite on the ground.
2. IK refinement. Per-frame mink IK moves the result toward the pelvis-local pose of both feet and both hands of the G1, position and orientation. A per-DOF posture cost holds the solution near step 1, so the step adjusts the pose and does not rearrange it.

The IK solves against the Lite Pro MJCF of robot-assets, and the robot layer draws the URDF of the same package. A new asset version therefore gives new motions, and retarget.py prints the pelvis shift it measured so that such a change is visible.

The file format

store.py writes dataset/ as one directory per layer: base/ for the source clip as published at 30 Hz, lite_pro/ for the 50 Hz retarget, and blueprints/ for the viewer layout. DATASET_CARD.md documents that store, and store.py documents every entity path.

The robot layer is what training and deployment read, and it is self-contained: its joints reset the robot, and its body poses are the goal, turned into the frame convention of the Lite links so a reward compares them against the robot directly. It stores no reached pose, because that is the forward kinematics of the joints beside it.

store.py is the only module that writes Rerun data, and it uses rerun-sdk alone. The layout on disk matches the motion store of MikuMotionTools, so its MotionStore reads these files.

The training repository, Lite-Motion-Tracking, still reads the older single-layer tracking/*.rrd. It needs a reader for this store before it can train on these motions.