AllenAct is an open-source framework for research in embodied AI, built and backed by the Allen Institute for AI (AI2). Published on GitHub with about 380 stars and still updated in May 2026, it provides a modular and flexible learning framework designed around the unique requirements of embodied agents.
AllenAct decouples tasks from environments, so researchers can implement a large variety of tasks in the same simulated world and swap algorithms without rewriting their setup. It offers first-class support for embodied environments including iTHOR, RoboTHOR, and Habitat, plus grid-worlds such as MiniGrid. The framework ships reproductions of state-of-the-art models, extensive documentation, tutorials, starter code, and pre-trained models, which lowers the cost of reproducing published results. A companion research paper describes the design, and the project targets experiments in visual navigation and related embodied tasks.
The project grew out of AI2’s work on embodied environments such as AI2-THOR, and its documentation site includes installation guides and tutorials that take a new user from setup to a first training run. Starter code and pre-trained models mean experiments can be reproduced and compared rather than rebuilt from scratch, which is the central frustration the framework was designed to remove from embodied AI research.
AllenAct is free and open source. There is no paid tier or hosted platform; researchers install the framework and run experiments on their own machines or clusters.
AllenAct is aimed at embodied AI researchers and graduate students who need a reproducible, modular codebase for training and evaluating agents in simulated households and 3D environments.