AllenAct is a modular and flexible learning framework designed with a focus on the unique requirements of Embodied-AI research. It provides first-class support for a growing collection of embodied environments, tasks and algorithms, provides reproductions of state-of-the-art models and includes extensive documentation, tutorials, start-up code, and pre-trained models.
AllenAct is built and backed by the Allen Institute for AI (AI2). AI2 is a non-profit institute with the mission to contribute to humanity through high-impact AI research and engineering.
- Support for multiple environments: Support for the iTHOR, RoboTHOR and Habitat embodied environments as well as for grid-worlds including MiniGrid.
- Task Abstraction: Tasks and environments are decoupled in AllenAct, enabling researchers to easily implement a large variety of tasks in the same environment.
- Algorithms: Support for a variety of on-policy algorithms including PPO, DD-PPO, A2C, Imitation Learning and DAgger as well as offline training such as offline IL.
- Sequential Algorithms: It is trivial to experiment with different sequences of training routines, which are often the key to successful policies.
- Simultaneous Losses: Easily combine various losses while training models (e.g. use an external self-supervised loss while optimizing a PPO loss).
- Multi-agent support: Support for multi-agent algorithms and tasks.
- Visualizations: Out of the box support to easily visualize first and third person views for agents as well as intermediate model tensors, integrated into Tensorboard.
- Pre-trained models: Code and models for a number of standard Embodied AI tasks.
- Tutorials: Start-up code and extensive tutorials to help ramp up to Embodied AI.
- First-class PyTorch support: One of the few RL frameworks to target PyTorch.
- Arbitrary action spaces: Supporting both discrete and continuous actions.
Environments | Tasks | Algorithms |
---|---|---|
iTHOR, RoboTHOR, Habitat, MiniGrid, OpenAI Gym | PointNav, ObjectNav, MiniGrid tasks, Gym Box2D tasks | A2C, PPO, DD-PPO, DAgger, Off-policy Imitation |
We welcome contributions from the greater community. If you would like to make such a contributions we recommend first submitting an issue describing your proposed improvement. Doing so can ensure we can validate your suggestions before you spend a great deal of time upon them. Improvements and bug fixes should be made via a pull request from your fork of the repository at https://github.com/allenai/allenact.
All code in this repository is subject to formatting, documentation, and type-annotation guidelines. For more details, please see the our contribution guidelines.
This work builds upon the pytorch-a2c-ppo-acktr library of Ilya Kostrikov and uses some data structures from FAIR's habitat-lab. We would like to thank Dustin Schwenk for his help for the public release of the framework.
AllenAct is MIT licensed, as found in the LICENSE file.
AllenAct is an open-source project built by members of the PRIOR research group at the Allen Institute for Artificial Intelligence (AI2).
If you use this work, please cite our paper:
@article{AllenAct,
author = {Luca Weihs and Jordi Salvador and Klemen Kotar and Unnat Jain and Kuo-Hao Zeng and Roozbeh Mottaghi and Aniruddha Kembhavi},
title = {AllenAct: A Framework for Embodied AI Research},
year = {2020},
journal = {arXiv},
}