Berkeley-Humanoids/Lite-Kimodo-Dataset

Lite Kimodo Dataset

This dataset holds the motions of the three Kimodo G1 sets, retargeted to the Berkeley Lite Pro humanoid at 50 Hz by Lite-Motion-Retargeting. Each motion is one Rerun recording, with the G1 clip that it comes from beside it. The Lite Motions Space plays all of them on one page.

Source	Set	Motions	Hours	A clip
kimodo	Kimodo G1 v1	9,490	10	4 s, one action
kimodolh2	Kimodo G1 long-horizon v2	9,947	54	about 20 s, 3 to 8 actions
kimodolh3	Kimodo G1 long-horizon v3	9,941	118	about 43 s, about 50 actions

Kimodo is not motion capture. NVIDIA's Kimodo model generated each clip from a text prompt, such as "a person throwing boxing punches". Then a tracking policy played the clip on a simulated Unitree G1. A motion is named <source>_<group>_episode_<index>. The group is the action that the prompt asks for, such as bowing_forward, or the family of a long-horizon prompt, such as combat_drill.

hf download Berkeley-Humanoids/Lite-Kimodo-Dataset --repo-type dataset \
    --include "*/kimodo/kimodo_bowing_forward_episode_00001.rrd" "assets/*" "blueprints/*" \
    --local-dir .

rerun */kimodo/kimodo_bowing_forward_episode_00001.rrd --asset assets/lite_pro.rrd \
    blueprints/lite_pro.rbl

The motions share the geometry of the robot, in assets/lite_pro.rrd. The geometry takes 9 MB, and a 4 s motion takes 1 MB, so each motion does not carry a copy. Pass the asset to rerun with --asset, and pass the blueprint last. The Rerun web viewer cannot load an asset, so use the Space to see a motion in a browser.

Layout

Each top directory is a layer, with a directory per source, and each file is a motion. This layout follows rerun/arkitscenes-rrd. All layers of one motion share an application_id and a recording_id, so the viewer opens any set of them as one recording.

Path	What it holds
base/<source>/<motion>.rrd	The source clip as published, at 50 Hz: the 29 G1 joints, and the world poses that these joints put the G1 bodies at. It records provenance, so you can derive the motion again without the upstream dataset.
lite_pro/<source>/<motion>.rrd	Everything the robot needs, at 50 Hz: the retargeted joints, the goal that its bodies track, and the joints that the source drives.
assets/lite_pro.rrd	The URDF geometry that draws the robot, for every motion.
blueprints/lite_pro.rbl	The viewer layout.
web/	The robot layer in the form that the motion browser plays: index.json.gz lists the motions, robot.json and robot.bin draw the robot, and each pack under packs/ holds float32 rows of motions end to end. src/lite_motion_retargeting/browser.py of Lite-Motion-Retargeting documents the format.

The robot layer is self-contained, and its joints reset the robot. /lite_pro/goal/ holds the goal that the robot tracks, in the frame convention of the Lite links, so a reward compares the goal with the robot directly. source_body_names records the G1 link that each goal came from. The layer stores no reached pose, because the reached pose is the forward kinematics of the joints beside it. The /lite_pro/tf chain draws the robot at that pose.

Both layers carry the pelvis, both feet, and both hands: /lite_pro/goal/ under the Lite link names, and /base/body/ under the G1 link names. A G1 clip drives 29 joints of Lite Pro, and active_joint_names lists them. The neck, the ankle yaw and the fingers have no G1 joint, so they hold 0. The metrics property records how well the retarget went.

Every array is in the world frame, in meters and radians, and each rotation is xyzw. Each velocity is the backward difference from the frame before, so frame 0 holds zero. Every column carries a frame index and a time in seconds. src/lite_motion_retargeting/store.py of Lite-Motion-Retargeting lists every entity path and property.

Source

Each set holds 10,000 episodes on the Hub, as LeRobot shards. The last shard of each set has no footer, and no reader can open it, so this dataset has fewer motions. Lite-Motion-Retargeting copies each G1 joint to its Lite joint, and refines the copy with a per-frame IK that matches the feet and the hands of the G1. Then it shifts the motion so that the soles stand on the floor. Lite-Motion-Retargeting documents every step. The result depends on the Lite Pro model, so a new version of Lite-Description gives new motions.

Apache-2.0, inherited from the three sets. The sets come from cagataydev. Their clips come from nvidia/Kimodo-G1-RP-v1, tracked on the G1 by cagataydev/protomotions-gtp-unitree-g1 with strands-labs/robots.