DexJoCo bimanual_assembly — 1000 generated demonstrations

1000 successful bimanual_assembly demonstrations for the DexJoCo benchmark, produced by SE(3)-warping the 100 official human teleoperation demos onto newly randomized object layouts and keeping the rollouts the simulator reports as successful (a DexMimicGen-style data-generation pipeline).

These are generated, not human-recorded. Every episode comes from warping one of the 100 official teleop demos to a fresh layout and rolling it out in MuJoCo. The on-disk format is byte-format identical to DexJoCo/DexJoCo-Datasets-Raw so it drops into the same tooling, but do not mistake it for the official raw data.

Task prompt: "Grasp the tray with the left hand and the peg with the right hand, then insert the peg into the hole."

Layout

dexjoco_raw_datasets/bimanual_assembly/<episode>/replay.zarr/
    data/{action (T,46) f64, action_rotvec (T,44) f64, state (T,1,61) f32, timestamp (T,) f64}
    meta/episode_ends
dexjoco_raw_datasets/bimanual_assembly/<episode>/videos/{ego,wrist_left,wrist_right}.mp4
metadata/episode_provenance.json          # per episode: source demo + step counts

state layout (61-d): tcp_pose(14) gripper_pose(32) socket_ori_pose(7) peg_ori_pose(7) table_delta_height(1). Videos are 640x640 H.264, 30 fps, one frame per timestep. Same zstd/BITSHUFFLE zarr codecs as the official assembly data.

Load an episode:

import zarr, numpy as np
g = zarr.open("dexjoco_raw_datasets/bimanual_assembly/<episode>/replay.zarr", mode="r")
state  = np.asarray(g["data"]["state"])[:, 0]     # (T, 61)
action = np.asarray(g["data"]["action"])          # (T, 46)

Convert to LeRobot / Zarr with the dexjoco-data-converter in the DexJoCo repo.

How it was generated

Per trial: sample a new layout (env.reset() — peg xy in a 5x5 cm box with +/-10 deg yaw, tray xy in a 10x10 cm box with +/-20 deg yaw, table height ~ U(0, 5 cm)); pick the source demo whose peg pose is nearest (top-3, random); warp each arm's end-effector trajectory by the per-subtask SE(3) delta and replay the 16-DoF Allegro hand joints verbatim; finish the insertion with a closed-loop servo onto the live socket-bottom geom; keep the rollout if the env reports success (peg/socket-bottom contact sustained 30 steps). Rendering is on during generation, so the written episode is exactly the rollout that was validated.

Yield: 1842 successes / 9785 trials = 18.8%. The residual failures are dominated by angular peg/hole misalignment introduced by the warp. From that pool, 1000 episodes were quota-sampled so the episode-length histogram matches the official demos'.

How closely it matches the official data

property	official (100 demos)	this set (1000 episodes)
episode length	median 550 steps / 18.3 s, mean 582, range 431-1069	median 559 / 18.6 s, mean 601, range 446-1170
length distribution	—	two-sample KS 0.114 vs official (crit 0.143 at a=0.05)
peg x / y layout	U(bounds)	100% in bounds, uniformity KS 0.035 / 0.038 (crit 0.043)
tray x / y layout	U(bounds)	100% in bounds, KS 0.018 / 0.026
peg / tray yaw, table height	U(bounds)	100% in bounds, KS 0.074 / 0.056 / 0.057 — mildly skewed
storage format	—	68/68 properties identical (zarr keys, shapes, dtypes, codecs, filters, fill_value, order, attrs, episode_ends, 30 Hz timestamps, action_rotvec == recompute(action), video names/codec/resolution/pix_fmt/fps/frame-count)
content	—	identical start pose, object block constant within an episode, unit-norm action quaternions, per-step motion + TCP tracking percentiles in the official range

Known deviations: the orientation axes and table height keep a mild skew from the success filter (angular misalignment is the dominant failure mode, so success is not orientation- neutral); commanded end-effector poses reach slightly outside the official command envelope (z to 1.726 vs 1.701 m); and 87 of the 100 official demos appear as sources, the most frequent contributing 68 episodes.

Provenance and reproduction

metadata/episode_provenance.json maps every episode to the official source demo it was warped from, with both step counts. The generator, the production driver (scripts/datagen_produce.py), the selection step (scripts/datagen_select_balanced.py) and the audit tooling (dexjoco/datagen/audit/) live on the datagen branch of the DexJoCo repo, and dexjoco/datagen/RESULTS.md records the full measurement log.

License

MIT, following the DexJoCo repository. The source demonstrations are the official DexJoCo teleoperation data; please cite the DexJoCo paper if you use this set.

@article{dexjoco,
  title  = {DexJoCo: A Unified Benchmark and Toolkit for Task-Oriented Dexterous Manipulation on MuJoCo},
  url    = {https://arxiv.org/abs/2605.16257},
  year   = {2026}
}