Episodes Preview SO-101 Visualizer
150 episodes · 15 fps · 2 cameras · 224×224 av1

so101-sim-pickplace-dr

150 episodes / 16,713 frames of successful SO-ARM100 pick-and-place demonstrations in MuJoCo, 15 fps, two RGB cameras.

Part of Sim2Cell, a from-scratch imitation-learning pipeline on a simulated industrial work-cell: a MuJoCo SO-ARM100 arm (5-DoF + jaw) picks a block off the floor and places it in a tray, recorded straight to LeRobotDataset v3 and used to train ACT from pixels.

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

This is the domain-randomized set: 150 successful expert episodes with the world re-randomized at every reset across two independent axes:

  • visual -- lighting pose/intensity/ambient, floor & tray colors, block hue, fixed-camera pose jitter
  • physics -- block mass (x0.5-2.0), contact friction (x0.7-1.4), actuator servo gain

Pristine model values are restored before each application, so randomizations never compound across resets. The scripted expert still succeeds under full randomization (validated before recording), so the data distribution is honest.

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 (same order)
observation.images.front 224x224 RGB video (AV1), fixed front camera
observation.images.wrist 224x224 RGB video (AV1), wrist-mounted camera

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("ahmedsohail2003/so101-sim-pickplace-dr")
# On Windows, add: video_backend="pyav"

Collection

Episodes are generated by a scripted 8-phase expert (damped-least-squares IK with a gripper-down constraint, block-yaw-aligned grasps) running headless in MuJoCo 3.10. Only episodes that end with the block verifiably inside the tray are saved -- failures are discarded at record time, so every episode is a successful demonstration.

The Sim2Cell dataset trilogy

Dataset Episodes Purpose
so101-sim-pickplace 100 Nominal scene -- baseline ACT training
so101-sim-pickplace-v2 160 Adds recovery demonstrations (perturbed first grasp -> verified retry)
so101-sim-pickplace-dr 150 Domain-randomized -- visual + physics variation for robustness training

Measured motivation for the trilogy (15 eval episodes per randomization level, identical protocol): an ACT policy trained on the nominal set reaches 13/20 (65%) from pixels (75% with temporal ensembling) but degrades under domain randomization -- 60% nominal down to 40% at full DR -- while the privileged scripted expert stays at 100% everywhere. Training the identical ACT recipe on the -dr set closes the gap and then some: 87-100% across every randomization level and 18/20 (90%) on nominal (act-so101-pickplace-dr). An earlier version of this paragraph claimed a 0% collapse for the nominal policy; that number was a measurement artifact (an eval-harness rendering bug fed the policy black frames) and is corrected here.

Provenance & license

Robot model vendored from MuJoCo Menagerie (trs_so_arm100, Apache-2.0), with TCP-site and wrist-camera patches. Dataset license: Apache-2.0.

Downloads last month
91

Models trained or fine-tuned on ahmedsohail2003/so101-sim-pickplace-dr