so101-sim-pickplace-dr2

210 episodes / 23,310 frames of successful SO-ARM100 pick-and-place demonstrations in MuJoCo under full domain randomization, with targeted coverage of an eval-identified failure pocket. 15 fps, two RGB cameras.

Task (language-labeled): "Pick up the red block and place it in the blue tray."

Why this dataset exists — closing a measured coverage hole

act-so101-pickplace-dr, trained on the 150-episode -dr set, scores 90% over 50 evaluation episodes — and all five of its failures cluster in one small near-field pocket (final block positions x∈[−0.036,−0.011], y∈[−0.209,−0.198]; the corresponding spawns fall at r ≈ 0.19–0.21 m, θ ≈ −11°…+1°; zero failures elsewhere). The scripted expert succeeds in that pocket under full randomization (verified), so the hole is data coverage, not task difficulty.

This set therefore samples ~30% of episodes from that pocket region (r∈[0.19,0.23], θ∈[−14°,+2°]) and the remaining ~70% from the full workspace, all under the same per-reset visual + physics randomization as -dr (lighting/colors/camera jitter; block mass ×0.5–2, friction ×0.7–1.4, servo gain). All 210 episodes are expert successes (0 discarded at record time).

Measured outcome (2026-07-28) — a documented negative result. The retrain (act-so101-pickplace-dr2, 75k steps, approximately epoch-matched at ~25.7 vs ~28.7) scored 44/50 (88%) vs ACT-DR's 45/50 (90%): the pocket oversampling closed the target pocket everywhere probed (12/12 on a fixed spawn grid where ACT-DR scores 5/12) but opened a new failure pocket at far reach (0/8 at r ≥ 0.259 m, where ACT-DR is 12/12) — consistent with oversampling inside a fixed step budget reallocating precision rather than adding it. Full forensic on the model card and in the Part A README.

Format

Codebase	LeRobotDataset v3.0 (LeRobot 0.6.0)
FPS	15
observation.state	6-dim joint positions (Rotation, Pitch, Elbow, Wrist_Pitch, Wrist_Roll, Jaw)
action	6-dim joint-position setpoints
observation.images.front / .wrist	224x224 RGB video (AV1)

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("ahmedsohail2003/so101-sim-pickplace-dr2")
# On Windows, add: video_backend="pyav"

The Sim2Cell dataset family

Dataset	Episodes	Purpose
so101-sim-pickplace	100	nominal baseline
so101-sim-pickplace-v2	160	+ recovery demos (documented negative result for chunked BC)
so101-sim-pickplace-dr	150	domain-randomized (ACT-DR: 90% nominal, 87–100% robust)
so101-sim-pickplace-dr2	210	DR + pocket oversample (closed the pocket, moved the hole — net 88%)

Provenance & license

Scripted-expert demonstrations (success-filtered), robot model from MuJoCo Menagerie (trs_so_arm100, Apache-2.0). Dataset license: Apache-2.0.