panasonicai/xarm7_sim_auto_shaver_box_v2

xArm7 sim shaver box: autopilot demos with varied layouts

65 successful episodes (418,460 frames, 60 fps) of the bimanual shaver-box packing task in the MuJoCo / robosuite sim of the xArm7 workcell (KKallidromitis/xarm-robotics, sim/). A scripted autopilot recorded them; only successful episodes are kept. The dataset is meant for co-training with the real teleop data panasonicai/shaver_v2_human. It uses the same cameras, the same 16-D state and action layout, and the same task string.

Task

The task has 4 stages:

1. The right arm puts the covered shaver in its slot.
2. The left arm puts the charger in its slot.
3. The right arm slides the paper booklet into the box.
4. The left arm closes the lid and the right arm tucks the flap.

The booklet is the bendable paper sheet (--booklet paper). The foam is elastic, with a coarse 25x37 surface grid (XARM_SIM_FOAM_GRID=25x37).

Layout variation

Each episode draws a new layout. Variation is wider than the upstream sim default, so a policy can't rely on one fixed habit.

Object	Range
Box (with tray, lid, flap)	±2 cm across the left arm's reach, ±0.5 cm along it (+x capped at +1.5 cm), yaw ±3°
Foam	±2 cm
Paper booklet	±2 cm on the foam, yaw ±20° about its rest pose
Shaver, charger	anywhere within reach of their own arm (reach ≤ 0.62 m, 65° cone), full yaw; each stays on its own side

Enable it with XARM_SIM_SHAVER_LAYOUT=wide (per-range overrides XARM_SIM_BOX_*, XARM_SIM_FOAM_XY_CM, XARM_SIM_BOOKLET_*, XARM_SIM_ITEM_*). Code: branch sim/shaver-layout (commit dc5a762) plus the headless recording commit (df6d52a). Autopilot success with these ranges was 78% in dry runs; failed attempts were discarded.

Each episode's drawn layout (seed, RNG state, box, foam and booklet offsets, item poses) is in layouts/episode_XXXXXX.json. merge_manifest.json maps episodes to the 5 recording shards.

Format

LeRobot v3.0, 60 fps.

Key	Shape	Content
observation.images.{side_1, side_2, wrist_left, wrist_right}	300x480 video	the calibrated ZED views, same roles as the real data
observation.state	16	[R joints 1-7 (rad), R gripper, L joints 1-7, L gripper]; measured joints, commanded gripper
action	16	same layout; commanded joints and gripper
sim_qpos, sim_qvel	364 / 361	full MuJoCo state (sim only)

• Gripper: recorded binary 0/1 (--gripper-binary), because the sim jaw only opens or closes fully. The real data is analog.
• Prompt: the same 4-stage task string as panasonicai/shaver_v2_human.
• Features: the real data has motor_current, gripper_load and camera_time instead of the sim_* keys. Training code that reads only images, state, action and prompt can mix the two.

Merging shards: tools/merge_shards.py

The script that built this dataset from the 5 recording shards. It wraps lerobot.datasets.aggregate.aggregate_datasets, which copies the videos without re-encoding, and renumbers the layout sidecars:

python tools/merge_shards.py --out-repo <user>/<name> --out-root <dir> \
    --shard local/shard_a=<dir_a> --shard local/shard_b=<dir_b> ...

To co-train with the real data without merging the two (their extra columns differ), the openpi fork in xarm-robotics (branch cotrain-mix) mixes sources at load time, with DataConfig.mix = ((repo, episode_fraction, seed), ...). Held-out eval episodes come only from the primary (real) dataset.