AllenAct is a modular and flexible Python framework for Embodied AI research, developed by the Allen Institute for AI (AI2). It provides researchers with tools to easily implement and reproduce state-of-the-art algorithms across various embodied environments and tasks, accelerating progress in the field.

How It Works

AllenAct is built on a modular design that decouples environments and tasks, allowing for flexible experimentation. It offers first-class support for PyTorch and a wide range of reinforcement learning algorithms, including on-policy methods like PPO and DD-PPO, imitation learning, and offline training. The framework's architecture facilitates combining multiple losses and experimenting with sequential training routines, crucial for developing effective embodied agents.

Quick Start & Requirements

• Install via pip: pip install allenact
• Requires PyTorch.
• Supports environments like iTHOR, RoboTHOR, Habitat, MiniGrid, and OpenAI Gym.
• Official Docs: https://allenact.org/
• Tutorials: https://allenact.org/tutorials/

Highlighted Details

• Supports multiple embodied environments (iTHOR, RoboTHOR, Habitat, MiniGrid) and tasks (PointNav, ObjectNav).
• Implements various RL algorithms including PPO, DD-PPO, A2C, DAgger, and offline IL.
• Features multi-agent support, arbitrary action spaces (discrete/continuous), and integrated Tensorboard visualizations.
• Provides pre-trained models and extensive tutorials for rapid onboarding.

Maintenance & Community

• Developed by the PRIOR research group at AI2.
• Contributions are welcomed via GitHub issues and pull requests.
• Code adheres to formatting, documentation, and type-annotation guidelines.

Licensing & Compatibility

• MIT License.
• Permissive for commercial use and integration with closed-source projects.

Limitations & Caveats

The framework is primarily focused on research and may require significant effort to adapt for production deployment. Specific hardware requirements for certain embodied environments are not detailed in the README.