test_pick_mustard_200_cartesian_480

Raw teleoperation recording from the grabette handheld device: SLAM-tracked end-effector motion plus the two gripper joint angles, as demonstrated. Nothing here is post-processed into a policy-specific representation.

At a glance

episodes	199
frames	38963
fps	50
duration	~13.0 min
camera	observation.images.cam0 at 480x360
codebase_version	v3.0

Channels

action (11D)

channels	meaning
dx, dy, dz	end-effector translation delta
dr6d_0 … dr6d_5	end-effector rotation delta, 6D rotation representation
proximal, distal	gripper joint angles, radians (0 = open, positive = closing)

observation.state (2D): proximal, distal — the gripper only. The end-effector pose is deliberately absent: it is SLAM-frame-dependent, so feeding it to a policy ties the model to one recording session's origin.

is_lost flags frames where SLAM tracking was lost. Filter or reject those episodes before training — the pose deltas are meaningless there.

Tasks

task	episodes
test_pick_mustard_200	199

The gripper channels are RAW angles

A position-controlled servo replaying a demonstrated gripper angle under-closes: the recorded angle is where the human's fingers sat while pressing the object, so reproducing it stops just short and grips nothing. Across these recordings the demonstrations use only 38–60% of the proximal range.

Train on this dataset directly and the policy inherits that problem. The fix is to re-express the two angles as a grasp shape plus a closure that can be commanded to 1.0 — "close all the way" — letting the object stop the fingers. See docs/grasp_projection.md and grabette_postprocess.grasp_projection_convert.

Loading

from lerobot.datasets import LeRobotDataset

ds = LeRobotDataset("SteveNguyen/test_pick_mustard_200_cartesian_480")

Resolved by the git tag matching codebase_version (v3.0), not by main.