# DemystifyActionSpace Dataset

## Overview

This dataset is collected for the paper  
**"Demystifying Action Space Design for Robotic Manipulation Policies"**.

The dataset contains teleoperated demonstrations for a set of tabletop manipulation tasks using AgileX robotic platforms. It is designed to support empirical studies on how different action space designs affect policy learning for robotic manipulation.

The dataset includes both single arm and bimanual tasks, as well as a small cross embodiment subset collected on a different robot platform.

## Tasks

### touch — Touch Cube
The robot reaches toward a cube and makes contact with it.

This task evaluates basic reaching and contact behavior.

### pick_cup — Pick Up Cup
The robot grasps a cup and lifts it from the table.

This task focuses on grasp acquisition and lifting.

### pick_place — Pick and Place Cup
The robot picks up a cup and places it at a target location.

This task evaluates sequential manipulation involving grasping, transport, and placement.

### bowl — Bimanual Cube Transfer
The robot uses two arms to transfer a cube from one gripper to the other.

This task evaluates bimanual coordination and object handoff.

## Cross Embodiment Dataset

To study cross embodiment generalization, we provide a small subset collected on a different robot platform.

### airbot(touch) — Touch Cube on AIRBOT
The same Touch Cube task performed using the AIRBOT platform.

The subset of **touch** and **airbot(touch)** can be used to evaluate policy transfer across robot embodiments.

## Data Format

Each trajectory contains:

- observations  
- RGB images  
- robot states  
- actions  

The dataset is stored in a trajectory based format compatible with robot learning pipelines.

## Intended Use

This dataset is intended for research on:

- imitation learning for robotic manipulation  
- action space design in robot policies  
- policy learning with teleoperated demonstrations  
- cross embodiment policy generalization