ur10e-cup

VR-teleoperated UR10e + Robotiq 2F episodes of a single task: "pick up the cup". Recorded on a real robot cell — there is no simulator for this setup, so every evaluation is a real-robot rollout.

This dataset was created using LeRobot (dataset codebase v3.0).

Dataset summary

robot	UR10e (6 DoF) + Robotiq 2F gripper
episodes / frames	81 / 49,779
fps	20 (native, stride 1)
duration	~41 min of teleoperation
tasks	1 — "pick up the cup"
cameras	observation.images.side, observation.images.wrist — 480×640, AV1
size	~1.1 GB
splits	train: 0:81

Features

key	dtype	shape	notes
observation.images.side	video	(480, 640, 3)	third-person camera
observation.images.wrist	video	(480, 640, 3)	wrist camera
observation.state	float32	(7,)	6 UR joint angles (rad) + gripper
action	float32	(7,)	absolute joint targets + gripper, i.e. state[t+1]

Joint order: shoulder_pan, shoulder_lift, elbow, wrist_1, wrist_2, wrist_3, gripper. The gripper channel is binary (1 = open). Actions are absolute joint targets, not deltas — this matches the SO-100/SO-101 joint-vector convention that lerobot/smolvla_base was pretrained on.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("khanhnd61/ur10e-cup")
sample = ds[0]
print(sample["task"], sample["observation.state"], sample["action"].shape)

Provenance

Raw HDF5 recordings → frame/state extraction → gripper-timing correction → LeRobot v3.0 conversion. The gripper fix is the reason this extract is preferred over the un-corrected one: in the raw episodes the fingers start moving ~2.0 s before obs_gripper flips, so the binary edge was re-timed to the measured mid-transition against the video. At 20 Hz the original lag would have been ~40 steps of "gripper says open, fingers are closing".

Known caveats

• wrist_2 is a dead channel. State/action index 4 spans [1.5702, 1.5716] with std ≈ 0.0008 rad — the joint never moves, and that 1.4 mrad spread is just 4-decimal rounding. Under MEAN_STD normalization this dimension is amplified into near-pure noise. Consider dropping it.
• Lossy source. Frames come from CRF-26 H.264 and were re-encoded to AV1 (~0.006 MAE round-trip). Joints are rounded to 4 decimals, gripper to 3.
• Single task, single scene. 81 episodes of one instruction — expect a policy that masters this cell rather than one that generalizes. Language conditioning is effectively unused.
• No held-out split. All 81 episodes are in train; hold some out yourself if you want offline evaluation.

Training note

lerobot/smolvla_base declares its cameras as observation.images.camera1/2/3, so training against this dataset needs a rename:

--rename_map='{"observation.images.side": "observation.images.camera1",
               "observation.images.wrist": "observation.images.camera2"}'

The same mapping applies at rollout time.