Rachata/Umi-Sim-Gripper_pick-tomato-can

Umi-Sim-Gripper — pick a tomato soup can into a box

Two arm-less SO-101 grippers, flown by hand in VR, picking a YCB tomato soup can and placing it in a box. Recorded in Isaac Sim: no robot arm, no hardware, no rig — the operator wears a headset and each controller is a floating gripper, which is what makes a demonstration cost a minute rather than a workspace.

Episodes	16
Frames	3,674 (122.5 s)
Episode length	179–268 frames (6.0–8.9 s)
Control rate	30 Hz
Cameras	2 wrist, 480×360, 170° fisheye
Format	LeRobot v3.0
Simulator	Isaac Sim 6.0 / Isaac Lab 3.0

Loading

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Rachata/Umi-Sim-Gripper_pick-tomato-can")
print(ds.num_episodes, ds.num_frames, ds.fps)   # 16 3674 30

Verified with lerobot 0.6.1: pulled from the hub, decoded, indexed.

Schema

observation.state and action are 16 wide — this is a free-flying gripper, so there are no joints to record beyond the jaw:

right_x right_y right_z right_qx right_qy right_qz right_qw right_jaw
left_x  left_y  left_z  left_qx  left_qy  left_qz  left_qw  left_jaw

Position in metres in the environment frame, orientation as a unit quaternion (x, y, z, w), and the two halves are the right and left controller in that order.

State and action are different quantities, deliberately. The state is where each gripper's grasp point actually is, read back out of the simulator, with the jaw as the finger joint actually sits (in metres, 0 open to −0.044 closed). The action is what the operator commanded that step — the same pose convention, but the jaw as the binary ±1 the wire carries (+1 open, −1 closed). Storing the command in both columns would teach a policy that the world always does what it is told; here the gap between them is the contact.

Position tracking error between the two averages 5 mm, which is the servo's, not noise.

observation.images.a_wrist_right and observation.images.b_wrist_left are the two wrist cameras, 480×360, a 170° Kannala-Brandt fisheye matching the LeRobot wrist lens. Each is mounted past the gripper housing rather than on the finger axis — a lens looking straight down the fingers has the gripper's own body between it and the object. At the start pose the can sits 3.6° and 5.0° off centre in the two views, with the fingertips in the bottom corners.

How it was recorded

WebXR page → WebSocket → a ZeroMQ bridge → Isaac Sim (the so101-scene project: scripts/floating_gripper.py, configs/vr_free_gripper_can.yaml). GRIP is a clutch that moves and turns a gripper, TRIGGER closes its jaw, right B/A start and end an episode. Each gripper is a velocity servo rather than a pose write, so contact forces stay bounded and the jaws can actually close on something instead of launching it.

Known limits

• Small. 16 episodes is enough to fine-tune against, not to train from scratch.
• The jaw action is binary, so action[7] and action[15] sit at +1 for ~82% of frames — open is the resting state, not a stuck column.
• 5–8% of frames in some episodes repeat the previous action exactly. Those are real operator pauses: with the clutch released the gripper holds its pose.
• Simulation only. Nothing here has been validated on a physical SO-101.

Provenance

Rebuilt from the original v2.1 upload: converted to the v3.0 layout, meta/stats.json computed over every frame, and three episodes of 5, 12 and 15 frames removed — mis-pressed buttons in which the gripper never moved, below ACT's chunk size and unusable as training samples. The surviving 16 are renumbered 0–15.