Bigenlight/carrot_in_pot_lighting_lerobot_v3

carrot_in_pot_lighting — UR7e teleoperation under added lighting (LeRobot v3.0)

Joint-space imitation-learning dataset with 65 human teleoperation episodes of a UR7e arm putting a plastic carrot into a pot, recorded from two 720p RGB cameras. This is a new, separate collection from Bigenlight/carrot_in_pot_lerobot_v3.

Task scene captured on 2026-09-18. The two visible softbox lights are the added-lighting setup used during this collection.

• 65 episodes · 20,472 frames · 30 fps · 11.37 min · about 299 MB
• Format: LeRobot v3.0 (lerobot 0.6.1), robot_type: "ur7e_gello"
• State/action: 7-D absolute joint representation (6 UR joints + gripper)
• Two RGB views: cam1 scene camera and cam2 wrist camera
• RGB only; no depth features
• Recorded on 2026-09-18 using the fixed stamp_s recorder
• No timestamp correction and no stale-tail frame removal were applied

Important difference from the earlier carrot dataset: added lighting

Additional task lighting was installed and used throughout this entire 2026-09-18 collection session. This is the main visual-domain difference from Bigenlight/carrot_in_pot_lerobot_v3, which was collected before the added lighting.

The light changes scene appearance, brightness, shadows, highlights, and potentially camera auto exposure. Treat this dataset as a distinct lighting condition when splitting, mixing, or evaluating policies. The lamp model, position, intensity, and color temperature were not instrumented, so this card intentionally makes no more specific photometric claim.

The earlier dataset contains 54 episodes with RGB and native depth and required a measured 0.895–0.900 s state-timestamp correction. This new dataset is RGB-only and was captured with the fixed recorder, so it is aligned directly from per-row ROS header stamps without that correction.

Recovery labels and source-name preservation

The original take folder name is preserved verbatim for every episode in meta/source_takes.json, under episodes[].take. Do not infer the take number from episode_index and do not rename these values. Take numbering restarted across collection batches, so the full timestamped name is the stable identifier.

The suffixes are operator annotations from collection:

• _recover: the first task attempt failed, and the operator recovered and tried again within the same continuous take.
• _recover_twice: the take contains two recoveries/retries after failed attempts.

There are eight _recover takes and one _recover_twice take:

episode	exact source take	label
17	take_06_20260918_202527_recover	recover
20	take_07_20260918_202637_recover	recover
22	take_08_20260918_195306_recover	recover
23	take_08_20260918_202709_recover	recover
26	take_09_20260918_202737_recover_twice	recover twice
59	take_41_20260918_200952_recover	recover
60	take_42_20260918_201017_recover	recover
62	take_44_20260918_201121_recover	recover
63	take_45_20260918_201147_recover	recover

The same lists and label definitions are stored machine-readably under take_name_labels in meta/source_takes.json, so recovery episodes can be selected later without parsing this card. These labels describe within-episode behavior; there is no per-step reward or failure column.

Setup and task

The physical setup is a Universal Robots UR7e, a passive GELLO leader arm, a Robotiq 2F-85 parallel gripper, and two Intel RealSense cameras. Collection used end-effector-delta teleoperation: the bridge converted leader-pose deltas into absolute UR joint targets through IK before recording. Therefore action is an absolute joint target even though teleoperation was performed in EEF space.

Task string, exactly as stored: Put carrot in pot.

Schema

feature	dtype	shape	meaning
observation.state	float32	(7,)	measured UR joints ur_q1..6 in radians + continuous grip_pos
action	float32	(7,)	commanded absolute joints cmd1..6 in radians + grip_cmd
observation.images.cam1	video	(720, 1280, 3)	scene RGB, AV1
observation.images.cam2	video	(720, 1280, 3)	wrist RGB, AV1

Standard LeRobot bookkeeping columns (timestamp, frame_index, episode_index, index, and task_index) are also present. Gripper convention is 0 = open, 1 = closed.

Episode lengths are 205–649 frames (6.8–21.6 s), with a median of 296 frames. The longest episode is the _recover_twice take. All 20,472 raw cam1 frames are retained.

Alignment and conversion

One episode is produced per raw take. cam1 is the master stream. The fixed recorder stores ROS header stamp_s values for camera and measured-joint rows; these capture instants are used directly for nearest-timestamp alignment. Command and gripper streams use their callback timestamps. No lag is subtracted and no frames are dropped.

The exact episode-to-take mapping, alignment method, per-episode frame count, timestamp-offset diagnostic, camera-frame reuse count, lighting condition, and recovery lists are in meta/source_takes.json.

Verification

The conversion was checked with the independent validator scripts/dataset/validate_takes_conversion.py, which does not import the converter:

• 43 passed, 0 failed, 0 warnings, 0 skipped — verdict PASS
• 65/65 episodes and 20,472/20,472 frames accounted for
• independently re-derived state/action: maximum absolute difference 0.0
• no NaN or infinity in state, action, or timestamps
• cam1 video: 390 sampled frames, correlation min 0.99931, mean 0.99964
• cam2 video: 390 sampled frames, correlation min 0.99950, mean 0.99978
• median |state[:6] - action[:6]|: 0.00470 rad
• per-episode median cam1↔joint timestamp offset: −0.03 to +0.02 ms
• no absolute local filesystem path under meta/

The machine-readable report is included as validation_2026-09-18.json.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Bigenlight/carrot_in_pot_lighting_lerobot_v3")
print(ds.meta.total_episodes, ds.meta.total_frames)  # 65, 20472

sample = ds[0]
sample["observation.state"]
sample["action"]
sample["observation.images.cam1"]
sample["observation.images.cam2"]

To select recovery episodes while preserving their exact identifiers:

import json
from huggingface_hub import hf_hub_download

p = hf_hub_download(
    "Bigenlight/carrot_in_pot_lighting_lerobot_v3",
    "meta/source_takes.json",
    repo_type="dataset",
)
source = json.load(open(p))
recover_names = set(source["take_name_labels"]["recover_takes"])
recover_names.update(source["take_name_labels"]["recover_twice_takes"])
recover_episodes = [
    ep for ep in source["episodes"] if ep["take"] in recover_names
]

Loading by repo ID requires the v3.0 tag, created after upload.

Related datasets

repository	distinction
Bigenlight/carrot_in_pot_lighting_lerobot_v3	this dataset — added lighting, fixed recorder, RGB-only, 65 episodes, recovery labels preserved
Bigenlight/carrot_in_pot_lerobot_v3	earlier real-robot carrot collection — pre-added-lighting condition, 54 episodes, RGB + depth, timestamp-corrected
Bigenlight/carrot_in_pot_raw	raw source for the earlier 54-episode dataset, not this collection
Bigenlight/carrot_in_pot_sim_lerobot_v3	simulated counterpart

Limitations

• One task, one robot setup, one operator, and one collection date.
• The added lighting condition is documented qualitatively; no calibrated illuminance or color temperature measurement is available.
• RGB only; no depth was recorded for this collection.
• Recovery is an episode-level name annotation, not a per-frame failure/reward label.
• The LeRobot conversion omits raw TCP pose, wrench, joint velocity/effort, and GELLO-leader streams.
• This release does not claim policy performance.

Citation

@misc{theo2026carrotinpotlighting,
  title        = {carrot_in_pot_lighting: UR7e teleoperation demonstrations under added lighting},
  author       = {Theo and {Bigenlight}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Bigenlight/carrot_in_pot_lighting_lerobot_v3}},
  note         = {LeRobot v3.0 dataset, 65 episodes, two RGB views, recovery labels preserved}
}