banana_in_pot_raw — raw HDF5 + MP4 teleop logs (UR7e, "put the right banana in the pot")

The raw, unprocessed recordings behind the banana_in_pot family: one folder per take, each with a native multi-rate HDF5 of every logged signal (UR joints with velocities and efforts, joint commands, TCP pose, 6-axis force/torque wrench, gripper, and the GELLO leader streams) plus the two raw camera MP4s. Use this if you want to resample differently, add features, or study the leader/follower relationship — otherwise start from the ready-to-train LeRobot datasets.

• 51 takes · 21,524 camera frames · 718 s (12 min) · 30 fps cameras
• Per take: vectors.h5 + cam1.mp4 + cam2.mp4
• Nothing resampled — each stream keeps its own t_rel_s clock and native rate.

Setup

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

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

Both stream 1280×720 (720p) @ 30 fps, yuv420p; raw files are cam1.mp4 / cam2.mp4 (MPEG-4 encoded). No depth or IR was recorded — despite the RealSense depth capability, only the RGB color stream was saved (no point cloud, no depth map). One camera is on a tripod (scene / third-person view), the other views the workspace; the two physical viewpoints (cam1 ↔ cam2) must be kept in order. An ArUco/AprilTag fiducial is present on the table.

Task

"put the right banana in the pot." The tabletop holds distractor objects — two bananas, an apple, carrots/peppers, and a slice of watermelon — plus a silver pot. The operator grasps the RIGHT banana (the target; everything else is a distractor) and places it in the pot. Success = the right banana ends up in the pot. All 51 takes are successful demonstrations.

HDF5 schema (vectors.h5)

Each group has its own t_rel_s (seconds, relative to take start) sampled at that stream's native rate. Field counts below are per-take examples (they vary with take length).

group	native rate	fields	units / meaning
cam1_frames	30 Hz	frame_idx, t_rel_s	index/time of each cam1 (D435) MP4 frame
cam2_frames	30 Hz	frame_idx, t_rel_s	index/time of each cam2 (D435if) MP4 frame
command	~56 Hz	cmd1..cmd6, t_rel_s	commanded absolute UR joint targets (rad)
ur_joint_states	~56 Hz	q1..q6, qd1..qd6, eff1..eff6, t_rel_s	UR7e follower joint positions (rad), velocities (rad/s), efforts (torque)
tcp_pose	~56 Hz	x,y,z, qw,qx,qy,qz, t_rel_s	TCP pose in robot base frame: position (m) + orientation quaternion
wrench	~56 Hz	fx,fy,fz, tx,ty,tz, t_rel_s	6-axis end-effector force (N) + torque (N·m)
gripper	~37 Hz	grip_pos, grip_cmd, gello_grip, t_rel_s	measured opening, commanded (binary open/close), leader gripper
gello_joint_states	~30 Hz	q1..q6, qd1..qd6, t_rel_s	GELLO leader joint positions (rad) + velocities (rad/s)
synchronized	—	(all keys, length 0)	EMPTY in every take — a scaffold group; ignore

Cameras: cam1.mp4 / cam2.mp4, 1280×720 RGB, 30 fps, MPEG-4. Use cam*_frames frame_idx / t_rel_s to align frames to the vector streams.

See DATA_DICTIONARY.md in this repo for the exhaustive per-field listing (every dataset key, dtype, and unit).

Notes:

• command (cmd1..6) are the joint targets sent to the UR7e; ur_joint_states.q* are the measured follower joints. gello_joint_states.q* are the leader joints — provided for completeness but not a valid inference input (the deployed robot cannot see the leader).
• tcp_pose is the recorded RTDE TCP (validated to be the flange pose; sub-mm / sub-0.2° agreement with forward kinematics at rest).
• The physical robot is a UR7e; the derived LeRobot datasets carry the legacy metadata label robot_type: "ur5e_gello".

How the derived datasets are built from this

dataset	contents
raw (this repo)	full multi-rate HDF5 + 2 MP4s — every signal above, native rates
LeRobot joint	observation.state(7)=ur_q1..6+grip_pos; action(7)=cmd1..6+grip_cmd; 2 AV1 videos
LeRobot EE	the above + observation.tcp_pose(7) + observation.wrench(6)

Both LeRobot datasets resample every stream onto the 30 fps camera grid (nearest timestamp on the cam1 clock), exclude the gello_* leader streams, and re-encode the videos to AV1. This raw release keeps everything at native rate so you can resample or add features yourself.

Usage

Read a take with h5py:

import h5py, cv2, numpy as np

take = "take_04_20260707_154857"
with h5py.File(f"{take}/vectors.h5", "r") as f:
    ur_q   = np.stack([f["ur_joint_states"][f"q{k+1}"][:] for k in range(6)], axis=1)  # (N56, 6) rad
    cmd    = np.stack([f["command"][f"cmd{k+1}"][:]       for k in range(6)], axis=1)  # (N56, 6) rad
    tcp    = np.stack([f["tcp_pose"][k][:] for k in "x y z qw qx qy qz".split()], 1)   # (N56, 7)
    wrench = np.stack([f["wrench"][k][:]   for k in "fx fy fz tx ty tz".split()], 1)   # (N56, 6)
    grip   = f["gripper"]["grip_pos"][:]                                               # (N37,)
    cam1_t = f["cam1_frames"]["t_rel_s"][:]                                            # (Ncam,) 30 Hz
    # each stream has its own f[group]["t_rel_s"] — align by nearest timestamp

cap = cv2.VideoCapture(f"{take}/cam1.mp4")   # 1280x720 RGB @ 30 fps

To go straight to training, use the ready LeRobot datasets instead (LeRobotDataset("Bigenlight/banana_in_pot_lerobot_v3")); a trained ACT policy is at Bigenlight/act_banana_in_pot.

Known quirks

• 1 take originally had NaN in grip_cmd (333 frames) — present in the raw log; repaired by forward/back-fill only in the derived LeRobot datasets.
• ~7 takes contain robot-stream dropouts (75–280 ms gaps in the ~56 Hz streams).
• grip_pos reaches ~0.898 (fully-open gripper) — physical, not an artifact.
• synchronized/ group is empty in all takes.

Related repositories (this family)

repo	contents
Bigenlight/banana_in_pot_lerobot_v3	main LeRobot joint-space dataset
Bigenlight/banana_in_pot_ee_lerobot_v3	EE variant (+ tcp_pose 7, wrench 6)
Bigenlight/banana_in_pot_raw	this — raw HDF5 + MP4
Bigenlight/act_banana_in_pot	trained ACT policy

The EE variant additionally feeds HIL-SERL-style end-effector-delta training (base-frame TCP displacement + gripper), derived from the tcp_pose stream above.

Limitations & intended use

• Single task, single scene layout, single operator; all demos are successes (no failure/recovery data).
• Streams are asynchronous (each has its own clock); you must resample/align them yourself — the synchronized/ group is empty.
• On fast motions tcp_pose carries a small timing-jitter offset relative to FK (an async logging artifact, not a kinematic error); at rest it agrees to sub-mm.
• Intended for research in imitation learning, teleoperation analysis, end-effector-space RL, and custom dataset construction.

Citation

@misc{theo2026bananainpotraw,
  title        = {banana_in_pot_raw: raw UR7e + GELLO teleoperation logs (HDF5 + MP4) for
                  "put the right banana in the pot"},
  author       = {Theo and {Bigenlight}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Bigenlight/banana_in_pot_raw}},
  note         = {51 takes, native multi-rate HDF5 + dual RGB MP4}
}

License: Apache-2.0.