Rachata/Umi-Sim-Arm_pick-place-crate

Umi-Sim-Arm — pick and place into a crate

Two SO-101 arms, teleoperated in VR, picking an object up and putting it in a crate. Recorded in Isaac Sim. This is the arm counterpart of Umi-Sim-Gripper_pick-tomato-can, which records the same operator and the same headset flying arm-less grippers instead — the two share a control protocol, a recorder and a camera lens, so they are directly comparable.

Episodes	2
Frames	636 (12.7 s)
Episode length	336 and 300 frames (6.7 s, 6.0 s)
Control rate	50 Hz
Cameras	2 wrist — right 480×360, left 320×240
Format	LeRobot v3.0
Simulator	Isaac Sim 6.0 / Isaac Lab 3.0

This is a two-episode dataset. It is a working sample of the recording path, not a training corpus. Use it to check a loader, a schema or a visualisation; do not expect a policy to come out of it.

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Rachata/Umi-Sim-Arm_pick-place-crate")
print(ds.num_episodes, ds.num_frames, ds.fps)   # 2 636 50

Schema

observation.state and action are 12 wide — the six SO-101 joints of each arm, right then left:

right_shoulder_pan right_shoulder_lift right_elbow_flex right_wrist_flex right_wrist_roll right_gripper
left_shoulder_pan  left_shoulder_lift  left_elbow_flex  left_wrist_flex  left_wrist_roll  left_gripper

Radians, except the gripper, which is its joint position.

State is measured, action is commanded. The state is the joints as they actually sit; the action is the joint targets the inverse kinematics set that step from the operator's hand pose — which is what a policy trained on this will be asked to produce.

observation.images.a_wrist_right (480×360) and observation.images.b_wrist_left (320×240) are the wrist cameras, a 170° fisheye matching the LeRobot wrist lens. The two differ in size because they were configured separately; the schema records each at its own resolution.

How it was recorded

WebXR page → WebSocket → a ZeroMQ bridge → Isaac Sim (the so101-scene project: scripts/run.py, configs/vr_teleop_crate.yaml). GRIP is a clutch that moves and turns an arm, TRIGGER closes its gripper, right B/A start and end an episode. The operator's hand pose drives task-space IK; the arms are position-controlled, so unlike the floating-gripper dataset the contact behaviour is the arm's own.

Known limits

• Two episodes. See the note above.
• No grasp actually happens. Measured across all 636 frames, the gripper joints move 0.001 and 0.008 rad — the operator flew the arms but never closed a gripper on the object. The task string above describes the scene, not what these two episodes accomplish.
• The motion is the left arm's (up to 2.5 rad of travel); the right arm stays within 0.24 rad.
• Simulation only. Nothing here has been validated on a physical SO-101.

Provenance

Recorded in v2.1 and rebuilt into the v3.0 layout with meta/stats.json computed over every frame. No episodes were dropped — both are well above the 16-frame floor. Verified by loading with lerobot 0.6.1.