robotfuel/so101_retargeted_umi

SO-101 handheld demonstrations, retargeted

48 pick-and-place demonstrations of a small box on a printed grid, recorded without a robot. A GoPro was mounted on a hand-held gripper mock-up in the UMI style; the resulting trajectories were retargeted into SO-ARM101 joint space and written in LeRobot format, so they load exactly like teleoperated episodes.

episodes	48
frames	11,284
fps	10
robot type	so101_follower
camera	wrist, 640×480
action / state	6 absolute joint positions, degrees

Episodes average 235 frames against 350 for teleoperated episodes of the same task — the hand moves faster than a teleoperator does.

Why it exists

To test whether cheap handheld demonstrations can stand in for expensive robot teleoperation. Sixteen of these episodes, added to a 16-episode teleoperated base, took a policy from 17% to 65% measured task success, where adding 16 more teleoperated episodes reached 70–75%.

The trained policies and the full comparison are at robotfuel/act_so101_t16b_u16.

Load it

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("robotfuel/so101_retargeted_umi")
print(ds[0]["observation.state"], ds[0]["action"])

What is and is not here

Frames are wrist-camera only — a downward view of the gripper, the box and the grid. No faces, no room, no bystanders.

The retargeting code is not included. These are its outputs.

Retargeting fidelity has not been measured directly against teleoperated trajectories; that comparison is outstanding, and it is the leading candidate to explain why scaling this data stops helping.