Allenact

AllenAct is an open source learning framework for Embodied AI research, built and backed by the Allen Institute for AI (AI2), a non profit research institute. It is designed for training and evaluating agents that perceive and act inside simulated environments, and it decouples tasks from environments so researchers can implement many different tasks in the same simulator without rewriting the surrounding code. The framework is for academic and industrial researchers working on navigation, manipulation and related embodied problems. It supports the iTHOR, RoboTHOR and Habitat environments as well as grid worlds such as MiniGrid, with plugins for ManipulaTHOR, BabyAI and Gym. On policy algorithms include PPO, a decentralized distributed form of PPO and A2C, alongside imitation learning, DAgger and offline imitation learning. Researchers can chain different training routines in sequence, combine several losses at once (for example a self supervised loss alongside a PPO loss), and run multi agent algorithms and tasks. TensorBoard integration visualizes first and third person agent views and intermediate model tensors. The project ships reproductions of published models, pretrained models, baseline configurations for object and point navigation, tutorials, how to guides and API documentation. AllenAct is free open source software with no paid tier, hosted service or pricing page. It is installed from the project documentation and GitHub repository, and an accompanying paper and citation are provided for academic use. It is a research framework rather than a finished product, so it suits teams comfortable configuring experiments in code.