GeorgiaTech/g1_bones_seed_sonic_129k_50hz

G1 BONES-SEED (SONIC-filtered, 129,785 clips, 50 Hz)

Unitree G1 reference motions retargeted from the BONES-SEED dataset, filtered with the public SONIC release keyword-exclusion filter (NVlabs/GR00T-WholeBodyControl) and exported to LAFAN1-style NPZ, then packed as WebDataset tar shards.

• Clips: 129,785 (SONIC keyword filter over seed_metadata_v004.csv; mirrors and non-neutral motions included). Selection SHA-256 in g1_bones_seed_sonic_selection.json.
• Rate: 50 Hz (resampled from the 120 Hz source).
• Bodies: 30 (G1_29DOF_DATASET_BODY_NAMES, canonical PhysX order; fixed rubber-hand links excluded). Joints: 29.
• Frame convention: every clip is anchored to a local origin (frame-0 root XY == (0, 0)); world positions carry no per-environment scene-grid offset, so the data is placement-independent (verified: same motion at different env slots produces byte-identical arrays).
• Quaternions: scalar-first wxyz in the NPZ (root_quat, body_quat_w).

Layout

shards/bones_seed_g1-XXXX.tar   # WebDataset shards, each member = <motion>.npz
shard_index.json                # shard -> members, byte sizes, per-shard sha256
g1_bones_seed_sonic_full_manifest.json    # LAFAN1-style manifest (129,785 trajectories)
g1_bones_seed_sonic_full_language.json    # per-motion language goal / category / metadata
g1_bones_seed_sonic_selection.json        # exact SONIC selection + filter keywords + hash

NPZ contents (per motion)

fps, qpos ([root_pos(3), root_quat(4,wxyz), joint_pos(29)]), qvel, root_pos, root_quat, root_lin_vel, root_ang_vel, joint_pos, joint_vel, body_pos_w, body_quat_w, body_lin_vel_w, body_ang_vel_w, joint_names, body_names.

Loading a shard

import tarfile, io, numpy as np
with tarfile.open("shards/bones_seed_g1-0000.tar") as tar:
    for m in tar.getmembers():
        d = np.load(io.BytesIO(tar.extractfile(m).read()))
        # d["body_pos_w"]: (T, 30, 3), d["joint_names"], d["body_names"], ...