This is the landing page for Taskography: Evaluating robot task planning over large 3D scene graphs, presented at CoRL2021. For a brief overview of the work, please refer to our project webpage.
To API. A simple API for sampling symbolic planning tasks in large-scale 3D scene graphs. Provides support for the following: (1) hierarchical-symbolic graph construction; (2) task sampling; (3) trajectory sampling; (4) procedurally generated environment wrappers.
To Envs. Official benchmark domains encoded in Planning Domain Definition Language (PDDL). PDDLGym enables gym-style interfacing for each planning domain and problem pair.
To Planners. A collection of performant, domain-agnostic learning-based planners, and the SCRUB and SEEK planners specifically designed to exploit 3D scene graph hierarchies.
To Baselines. A package interfacing all satisficing and optimal symbolic planners benchmarked in Taskography with Python. Planners are automatically installed on first use.
To Dev.
An in-house repository that contains the official benchmark results in .json file format, along with original scripts used to generate all Taskography planning domains.
Note: The majority of this code has been repurposed in Taskography-API.
All repositories are offered under the MIT License agreement. If you find Taskography useful, please consider citing our work:
@inproceedings{agia2022taskography,
title={Taskography: Evaluating robot task planning over large 3D scene graphs},
author={Agia, Christopher and Jatavallabhula, {Krishna Murthy} and Khodeir, Mohamed and Miksik, Ondrej and Vineet, Vibhav and Mukadam, Mustafa and Paull, Liam and Shkurti, Florian},
booktitle={Conference on Robot Learning},
pages={46--58},
year={2022},
organization={PMLR}
}