Dataset Viewer
Auto-converted to Parquet Duplicate
format_version
int64
key
dict
traj_info
dict
1
{"source":{"persist_id":"bones_seed_sonic_full_129785@e714bbff"},"manifest":"/mnt/storage/fwu91/bone(...TRUNCATED)
{"capacity":47491234,"written":47491234,"start_index":[0,184,530,1027,1759,2229,2738,3185,3317,3713,(...TRUNCATED)

Reference arrays: bones_seed_sonic_full_129785@e714bbff

Training-shaped reference arrays for humanoid motion tracking: 129,785 trajectories, 47,491,234 transitions, 49.4 GB. Built from GeorgiaTech/g1_bones_seed_sonic_129k_50hz.

This is a derived artifact, not a source dataset. It holds exactly the tensors an imitation environment consumes at run time, in their final layout, so a consumer memory-maps them instead of rebuilding a Zarr and a replay buffer. The authority on its identity is reference_arrays_manifest.json, which also carries the full trajectory table (start_index / end_index / ordered_traj_list) — without it the arrays are unloadable.

Contents

array shape dtype quaternion order
anchor_pos_w (47491234, 3) float32 -
anchor_quat_w (47491234, 4) float32 xyzw
body_ang_vel_w (47491234, 14, 3) float32 -
body_lin_vel_w (47491234, 14, 3) float32 -
body_pos_w (47491234, 14, 3) float32 -
body_quat_w (47491234, 14, 4) float32 wxyz
qpos (47491234, 36) float32 -
qvel (47491234, 35) float32 -

Retained bodies (order is the column layout, not a set): pelvis, left_hip_roll_link, left_knee_link, left_ankle_roll_link, right_hip_roll_link, right_knee_link, right_ankle_roll_link, torso_link, left_shoulder_roll_link, left_elbow_link, left_wrist_yaw_link, right_shoulder_roll_link, right_elbow_link, right_wrist_yaw_link

Baked anchor body: pelvis. Any other anchor within the retained set is derived at load time from the body arrays, so this artifact serves those too.

joint_pos and joint_vel are deliberately absent: they are qpos[:, 7:] and qvel[:, 6:].

Provenance

  • source manifest SHA-256: eb0ad052afe72fb6228f4be9d52132c9cb9a52ac9c561751e49e6ca31346e688
  • joints (29): recorded in the sidecar
  • full dataset body order (30): recorded in the sidecar

Use

python -m imitation_experiments.data.publish_reference_arrays fetch \
  --repo_id <this repo> --dest_dir /path/on/compute/node \
  --persist_id bones_seed_sonic_full_129785@e714bbff \
  --expected_motions 129785 \
  --expected_transitions 47491234

Then point the environment at it:

env.data.manifest=null
env.data.reference_arrays_dir=/path/on/compute/node
env.data.runtime_cache_device=cpu
env.data.macro_cache_device=cuda:0

The fetch validates the sidecar and every array's size against it. It does not re-verify contents against the source NPZs; do that at build time with build_reference_arrays.py --verify_load.

Downloads last month
-