Allen Institute for AI open-sources AllenAct, a modular and flexible learning framework designed to focus on the unique requirements of Embodied-AI research. Allen Institute (AI2) is a non-profit institute whose mission is to contribute to humanity through high-impact AI research and engineering.
Embodied AI, the AI subdomain concerning systems that learn to complete tasks through environmental interactions, has experienced substantial growth. AllenAct provides support for a growing collection of embodied environments, tasks, and algorithms, offers reproductions for state-of-the-art models, and includes extensive documentation, tutorials, start-up code, and pre-trained models.
Features:
- It supports multiple environments including iTHOR, RoboTHOR, and Habitat embodied environments.
- Using AllenAct, researchers can quickly implement a large variety of tasks in the same domain since tasks and environments are decoupled in AllenAct.
- It supports a variety of on-policy algorithms, including PPO, DD-PPO, A2C, Imitation Learning, and DAgger and offline training such as offline IL.
- It is trivial to experiment with different sequences of training routines, which are often the key to successful policies.
- It easily combines various losses while training models (e.g., use an external self-supervised loss while optimizing a PPO loss).
- It brings out of the box support to easily visualize first and third-person views for agents and intermediate model tensors, integrated into Tensorboard.
- It consists of various pre-trained models that can code and set a number of standard Embodied AI tasks.
Github: https://github.com/allenai/allenact
Paper: https://arxiv.org/abs/2008.12760