TreeePlanter/mujocothor_dataset

Dataset Card for MujocoThor Dataset

Dataset Summary

The MujocoThor Dataset is a collection of robot manipulation datasets from the MujocoThor simulation environment. It contains robot manipulation episodes with RGB images, robot states, and actions. The dataset is designed for training and evaluating robot manipulation policies in a simulated environment.

Supported Tasks and Leaderboards

• Robot Manipulation: The dataset supports training and evaluation of robot manipulation policies.
• Pick and Place: Contains episodes for pick and place tasks.
• Object Interaction: Includes various object manipulation scenarios.

Languages

The dataset is in English, with no natural language processing components.

Dataset Structure

Data Instances

Each instance is an episode containing:

• A sequence of robot states and actions
• RGB images from the camera
• Task metadata

Data Fields

• Observations:

  • image: RGB images from the left camera (720x1280x3)
  • state: Robot joint positions (7-dimensional)

• Actions: Robot end-effector pose from arm base (x, y, z, qw, qx, qy, qz)

• Metadata:

  • Task information
  • Episode information

Data Splits

The dataset is split into:

• Training set: 133 episodes

Dataset Creation

Source Data

Initial Data Collection and Normalization

The data was collected using the MujocoTHOR framework, which provides a realistic simulation environment for robot manipulation tasks. The framework simulates a Franka Emika Panda robot arm in a physics-based environment.

Who are the source language producers?

The data was generated using the MujocoTHOR framework developed by the Allen Institute for AI (AI2).

Annotations

Annotation process

The data is automatically generated through simulation, with no manual annotation required.

Who are the annotators?

N/A - Data is automatically generated.

Personal and Sensitive Information

The dataset does not contain any personal or sensitive information.

Additional Information

Dataset Curators

Allen Institute for AI (AI2)

Licensing Information

This dataset is licensed under the MIT License.

Contributions

@TreeePlanter.

Usage

import tensorflow_datasets as tfds

# Load a specific sub-dataset
dataset = tfds.load('mujocothor_dataset/franka_pickup', split='train')

# Iterate through episodes
for episode in dataset:
    for step in episode['steps']:
        image = step['observation']['image']
        state = step['observation']['state']
        action = step['action']

Available Sub-datasets

The repository includes the following sub-datasets:

• franka_pickup: Franka robot pickup task episodes
• (Future) franka_place: Franka robot place task episodes
• (Future) franka_push: Franka robot push task episodes
• (Future) Other task types...