so101_sim_pick_place

50 language-conditioned pick-and-place demonstrations on a simulated SO-101 arm, recorded in MuJoCo by the verified scripted skills of the so101-workcell project. Each episode places one coloured cube into one of two bins on a randomized layout. LeRobot v3 format, 10 Hz, two cameras.

episodes	50
frames	4,250 (85 per episode)
fps	10
cameras	observation.images.front (480x640), observation.images.wrist (480x640)
observation.state	6 joint positions in radians (shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper)
action	6 joint position targets in radians for the next control step
tasks	"put / move / place the {red,green,blue} cube in(to) the {left,right} bin"

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("saibala29/so101_sim_pick_place")
item = ds[0]  # observation.state, action, observation.images.front, observation.images.wrist, task

How it was made

python -m so101_workcell.hub.lerobot_export --episodes 50 --seed0 1000 from the project above. Scripted skills plan with inverse kinematics against local MuJoCo physics, verify every grasp and placement from simulator state, and the resulting 10 Hz joint-target stream is what a real SO-101 receives. Intended as simulated pre-training data to be mixed with real-arm teleoperation episodes.

Related

• Policy trained in the same simulator: https://huggingface.co/saibala29/so101-lift-sac
• Code, results and design notes: https://github.com/saibala2905/so101-workcell