Phongwit/crane_pick_place

CRANE+ V2 pick-and-place (Gazebo)

Simulated demonstrations of a CRANE+ V2 arm picking up a box from a table and placing it on the roof of a toy car. In each episode, the box and the car start at random positions. A scripted MoveIt expert performs the task, and the dataset records only episodes where the box ended up on the roof.

Code: github.com/bamboo51/Arm-Robot-Picking

Contents

Format	LeRobot dataset, 10 fps, images stored as MP4 video
Episodes	51 (training split only)
Task string	put the box on the car
Robot	CRANE+ V2 (4 arm joints and 1 gripper joint)
Simulator	Gazebo (Ignition) with ROS 2 Humble and MoveIt 2

Features

Key	Shape	Description
observation.images.top	240×320×3	camera above the table
observation.images.left	240×320×3	camera on the left side
observation.images.right	240×320×3	camera on the right side
observation.state	5	current joint angles (rad): crane_plus_joint1–joint4, crane_plus_joint_hand
action	5	joint angles 0.5 s (5 frames) later, in the same order

How the data was made

1. Reachable workspace. The link lengths and joint limits come from the URDF. Forward and inverse kinematics on a 2 mm (r, z) grid give the distances from the base at which the arm can grasp with the gripper pointing down. The result is a radius r of 0.080–0.124 m from the base's rotation axis and an angle θ of −130° to 130°.
2. Scenes. Each scene places the box and the car at random inside that range. The box center and the roof center are at least 0.11 m apart, and at least 60° apart around the base.
3. Expert. A MoveIt program adapted from RT Corporation's crane_plus examples carries out the task. It opens the gripper, moves above the box, descends, grasps, lifts, turns toward the car, reaches out, releases, and retreats.
4. Recording. rosbag2 (mcap) records /joint_states and the three cameras. Each recording is resampled to 10 Hz.
5. Success filter. After each episode, the box position is read from Gazebo. The episode is kept only if the box is within 3 cm of the roof center and within ±1 cm of the roof height.
6. Action target and idle frames. Each frame's action is the joint state 5 frames later. Frames in which nothing moves during the next 0.5 s are dropped. If they are kept, the policy learns to stay at the start pose.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("Phongwit/crane_pick_place")
frame = ds[0]
print(frame["observation.state"], frame["action"])

Limitations

• Simulation only. All images are rendered by Gazebo and have not been checked on a real robot.
• Few episodes for the workspace. 51 episodes over a 260° range give few examples per box angle, which limits how well a policy generalizes to new positions.
• Separate test split. The 48 test episodes are not part of this dataset. They are stored as raw episode.npz files.