Myungkyu/object_retrieval

instance_retrieval

Real-robot teleoperation demonstrations of the instance_retrieval task on a single-arm Franka Research 3 cell, released in four LeRobot layouts. Every layout is a conversion of the same 80 raw episodes (36,239 frames at 10 Hz); the layouts differ only in the LeRobot codebase version and in the action representation.

directory	LeRobot version	action (action)	consumer
lerobot_v21_abs_joint/	v2.1	8-D absolute joint targets + gripper	RLDX-1 loader (episodes.jsonl, meta/modality.json)
lerobot_v21_delta_eef/	v2.1	7-D end-effector velocity command + gripper	RLDX-1 loader
lerobot_v30_abs_joint/	v3.0	8-D absolute joint targets + gripper	lerobot ≥ 0.5 (LeRobotDataset), e.g. π0.5
lerobot_v30_delta_eef/	v3.0	7-D end-effector velocity command + gripper	lerobot ≥ 0.5

The v3.0 layouts were derived from the v2.1 layouts with the official lerobot converter (convert_dataset_v21_to_v30.py), so the four directories carry identical frames, states and videos.

Task

Collect every object that matches those shown inside the reference tray, place them into the collection box, and close the lid.

Every frame carries the task instruction in the task field (one task per dataset) and a subtask string column that holds the same instruction (reserved for subtask-level labels).

Raw data and conversion

Source: DROID-style teleoperation captures (one teleoperation.h5 per episode, HDF5) recorded with a Meta Quest controller driving the arm through a cartesian-velocity controller at a nominal 10 Hz. Each capture stores, per control step, the measured robot state (joint_positions, gripper_position, cartesian_position, joint velocities and torques), the commanded actions in several spaces (joint_position, cartesian_velocity, cartesian_position, gripper_position, …), controller flags, per-camera timestamps and intrinsics, and the camera streams as embedded mp4 bytes (two ZED stereo cameras, left and right eye each, 1280×720).

Conversion rules applied identically to all four layouts:

field	source in the capture	value
observation.state (8)	/observation/robot_state/joint_positions (7), gripper_position (1)	measured joint angles (rad) + measured gripper closedness (0 = open)
action (abs_joint, 8)	/action/joint_position (7), /action/gripper_position (1)	commanded absolute joint targets (rad) + commanded gripper closedness in [0, 1]
action (delta_eef, 7)	/action/cartesian_velocity (6), /action/gripper_position (1)	commanded cartesian velocity (vx vy vz wx wy wz, controller units in [-1, 1]) + commanded gripper closedness
observation.image.exterior_camera	exterior ZED 2i, left eye	448×252 RGB video (AV1, libsvtav1 crf 30), 10 fps
observation.image.wrist_camera	wrist ZED Mini, left eye	448×252 RGB video (AV1), 10 fps
timestamp, frame_index, episode_index, index, task_index	control-step index	standard LeRobot bookkeeping

Only the left eye of each stereo pair is kept; frames are resized from 1280×720 to 448×252 (same aspect ratio, no crop). The gripper is a Robotiq 2F; the state vector is the same in both action layouts, only the action head differs. The v2.1 layouts include meta/modality.json (state arm 0:7, gripper 7:8; action arm 0:7 / gripper 7:8 for abs_joint, eef_velocity 0:6 / gripper 6:7 for delta_eef) and meta/stats.json with the quantile statistics used for normalisation.

Statistics

	value
episodes	80
frames	36,239
fps	10
episode length	323–581 steps (mean 453)
cameras	2 (exterior, wrist), 448×252
robot type	franka_panda (Franka Research 3 arm, Robotiq 2F gripper)

Loading

# v3.0 layouts (lerobot >= 0.5)
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("Myungkyu/object_retrieval", root="<local copy>/lerobot_v30_abs_joint")

# any layout, plain download
huggingface-cli download Myungkyu/object_retrieval --repo-type dataset --include "lerobot_v21_abs_joint/*" --local-dir .

Companion checkpoints trained on these layouts: Myungkyu/hiwrld-baseline-ckpts-real-robot.