Bigenlight/banana_in_pot_lerobot_v3

banana_in_pot — UR7e teleop, "put the right banana in the pot" (LeRobot v3.0)

Joint-space imitation-learning dataset: 51 human teleoperation demos of a UR7e arm picking a target banana out of a cluttered scene and dropping it into a pot, with two 720p camera views. Ready to train ACT (or any LeRobot policy) out of the box.

• 51 episodes · 21,524 frames · 30 fps · ~11.96 min · ~483 MB
• Format: LeRobot v3.0 (lerobot 0.6.1)
• Action / state space: 7-D absolute joint (6 UR joints + gripper)
• Every demo is a success.

Setup

Collected on a Universal Robots UR7e — a 6-DOF collaborative arm (joint values in radians) — driven by a GELLO low-cost 3D-printed leader arm for kinesthetic teleoperation. The operator moves the GELLO leader; its joint positions are mapped to UR7e joint commands. Two Intel RealSense cameras record the scene as RGB video only:

• Camera 1 — Intel RealSense D435
• Camera 2 — Intel RealSense D435if (a D435 variant)

Both stream 1280×720 (720p) @ 30 fps, color format yuv420p. One camera is on a tripod (scene / third-person view), the other views the workspace. No depth or IR streams were recorded — despite the RealSense depth capability, only the RGB color stream was saved (no point cloud, no depth map in this release). The two physical viewpoints must be preserved at deploy time — cam1 ↔ cam2 order matters. An ArUco/AprilTag fiducial is present on the table.

Task

"put the right banana in the pot." The tabletop holds several distractor objects — two bananas, an apple, carrots/peppers, and a slice of watermelon — plus a silver pot. The operator must grasp the RIGHT banana (the target; the left banana and all other produce are distractors) and place it inside the pot.

• Success criterion: the right banana ends up inside the silver pot.
• Distractors: left banana, apple, carrots/peppers, watermelon slice.

Schema

feature	dtype	shape	meaning
observation.state	float32	(7,)	UR7e measured joints ur_q1..ur_q6 (rad) + grip_pos (gripper opening)
action	float32	(7,)	commanded absolute joint targets cmd1..cmd6 (rad) + grip_cmd
observation.images.cam1	video (AV1)	(720, 1280, 3)	viewpoint 1 (RGB, HWC uint8)
observation.images.cam2	video (AV1)	(720, 1280, 3)	viewpoint 2 (RGB, HWC uint8)

Plus the standard LeRobot bookkeeping columns (timestamp, frame_index, episode_index, index, task_index). action is an absolute joint-target command (not a delta); the gripper command grip_cmd is effectively binary (open/close).

Note: the LeRobot metadata records robot_type: "ur5e_gello" — a legacy label kept for reproducibility. The physical robot is a UR7e.

Hardware & Collection

• Robot: Universal Robots UR7e, 6-DOF, joints in radians. Inference uses the UR7e follower only.
• Teleoperation: GELLO leader arm (kinesthetic). The gello_* leader streams are intentionally excluded from this LeRobot dataset — the robot cannot observe the leader at inference time.
• Cameras: Intel RealSense D435 (cam1) + D435if (cam2), RGB only, 1280×720 @ 30 fps, yuv420p; raw cam1.mp4/cam2.mp4 were MPEG-4, re-encoded to AV1 (crf 30) here. No depth/IR.

Source stream rates (native, before resampling)

stream	native rate	in this dataset
cameras (cam1, cam2)	30 Hz	kept as the master 30 fps clock
command (UR joint targets)	~56 Hz	→ action[0:6], resampled to 30 fps
ur_joint_states	~56 Hz	→ observation.state[0:6], resampled to 30 fps
tcp_pose	~56 Hz	(not in this dataset — see EE variant)
wrench (6-axis F/T)	~56 Hz	(not in this dataset — see EE variant)
gripper	~37 Hz	→ grip_pos / grip_cmd, resampled to 30 fps
gello_joint_states	~30 Hz	excluded (leader-only)

All included robot streams are resampled onto the 30 fps camera grid by nearest-timestamp lookup on the cam1 frame timeline. cam2 is aligned to cam1 by nearest camera timestamp (handles the ±1–2 frame count mismatch seen in some takes).

• Scale: 51 episodes / 21,524 frames / 718 s (12 min) @ 30 fps.
• Episode set: raw take numbers 1–36 with gaps {9, 23, 30, 35} by design; re-takes fill the count back to 51 folders. All 51 episodes are successful demonstrations.

Known quirks (data characteristics, not conversion bugs)

• 1 take originally had NaN in grip_cmd (333 frames) — repaired by forward/back-fill during conversion.
• ~7 takes contain raw robot-stream dropouts (75–280 ms gaps), giving a cam↔robot nearest-match error up to ~40 ms (≈2 control periods) at those frames. Mostly harmless for ACT.
• grip_pos peaks at 0.898 (~6.3% of frames > 0.6) = fully-open gripper — physical, not an artifact.

QA: the conversion was independently validated (schema/v3.0 compliance, 4-way frame-count reconciliation, per-frame alignment re-derived to 1e-7, 0 NaN/Inf in stored arrays, video integrity corr 0.9995 vs raw, and a real lerobot-train smoke test).

Usage

Load with the LeRobot API:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Bigenlight/banana_in_pot_lerobot_v3")
print(ds.meta.total_episodes, ds.meta.total_frames)   # 51, 21524

sample = ds[0]
sample["observation.state"]          # (7,) float32
sample["action"]                     # (7,) float32
sample["observation.images.cam1"]    # (3, 720, 1280) float32 CHW

Train ACT (RTX 3060 12 GB reference; 720p resized on-the-fly to 360×640):

lerobot-train \
  --dataset.repo_id=Bigenlight/banana_in_pot_lerobot_v3 \
  --policy.type=act --policy.device=cuda \
  --batch_size=8 --steps=50000 \
  --output_dir=outputs/act_banana

A pretrained ACT checkpoint for this dataset is published at Bigenlight/act_banana_in_pot (chunk_size 100, ResNet18 + VAE). Reported open-loop held-out L1 ≈ 0.0235 rad (~1.34°), gripper accuracy 99.2%, with no train/held-out gap (no overfitting).

Related repositories (this family)

repo	contents
Bigenlight/banana_in_pot_lerobot_v3	this — main LeRobot joint-space dataset
Bigenlight/banana_in_pot_ee_lerobot_v3	EE variant: adds observation.tcp_pose (7) + observation.wrench (6)
Bigenlight/banana_in_pot_raw	raw HDF5 + MP4 (all original signals, incl. GELLO leader)
Bigenlight/act_banana_in_pot	trained ACT policy

The EE variant additionally supports HIL-SERL-style end-effector-delta training (base-frame TCP displacement + gripper).

Limitations & intended use

• Single task, single scene layout, single operator — expect limited generalization across novel object arrangements or lighting.
• All demonstrations are successes — no failure/recovery data; not suited as-is for methods that need negative examples.
• Offline metrics are strong, but closed-loop performance on a real UR7e is not guaranteed and should be validated on hardware.
• Intended for research in imitation learning and robot manipulation (ACT and related LeRobot policies).

Citation

@misc{theo2026bananainpot,
  title        = {banana_in_pot: UR7e teleoperation demonstrations for
                  "put the right banana in the pot"},
  author       = {Theo and {Bigenlight}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Bigenlight/banana_in_pot_lerobot_v3}},
  note         = {LeRobot v3.0 dataset, 51 episodes}
}

License: Apache-2.0.