---
license: cc-by-4.0
task_categories:
- robotics
language:
- en
tags:
- rlds
- ur3
size_categories:
- n<1K
---
This repo contains a robotic dataset of 300 episodes across three manipulation tasks, stored in RLDS fromat. The data was collected on `UR3 robot`, using MoCap-based teleoperation system with `force feedback`.
For all tasks, state observation consists of:
- An RGB image (128 × 128) captured from a wrist-mounted camera;  
- An RGB-D image (256 × 256) from a side-mounted RealSense D455 camera; 
- The robot’s proprioceptive state, including end-effector pose and joint angles;
- Gripper position and force sensor readings, standardized to the range [0, 1].

The action space is defined as a 7-dimensional vector, comprising:
- A 6-dimensional delta displacement for joint angles,
- 1-dimensional delta value for gripper position, standardized to the range [-1, 1]. 

The frequency of data collection was fixed at 10Hz.

The selected objects for manipulation included a `shampoo bottle`, a `tomato`, and a `tube of toothpaste`. 
Each trajectory starts from the robot's home position and ends when the robot successfully grasps the object and lifts it approximately 10 cm above the table surface.

---
license: cc-by-4.0
---