Summary

Air Learning provides a cross-domain infrastructure for studying and evaluating reinforcement learning (RL) algorithms for autonomous aerial robots. It targets researchers in robotics, RL, and systems architecture, bridging disparate fields to holistically address challenges like data requirements, energy, and onboard compute. The project aims to accelerate the development and deployment of robust aerial AI by enabling realistic simulation and evaluation.

How It Works

This framework integrates Microsoft AirSim, OpenAI Gym, Stable-Baselines, and Unreal Engine for aerial robot RL. It features a photorealistic environment generator (Unreal Engine/AirSim) with an energy model, exposed via an OpenAI Gym interface for integration with RL algorithms like DQN and PPO. A key innovation is the Hardware-in-the-Loop (HIL) methodology, simulating the environment on a desktop while running the RL agent on an embedded system. This allows evaluation across different onboard compute platforms and quantification of computational needs. Beyond success rates, Air Learning includes crucial "Quality of Flight" metrics: energy per mission, flight time, and distance traveled, important for resource-constrained UAVs.

Quick Start & Requirements

Tested on Windows 10; Linux support is pending. Installation involves two components: environment generator and RL training. Instructions are available via a linked repository or the air-learning-ue4 sub-module. Key dependencies include Unreal Engine, Microsoft AirSim, OpenAI Gym, and Stable-Baselines.

Highlighted Details

• Photorealistic, configurable, randomized environment generation (Unreal Engine/AirSim).
• Integrated energy model for aerial robots.
• OpenAI Gym interface for broad RL algorithm compatibility.
• Hardware-in-the-loop (HIL) simulation for evaluating performance on diverse onboard compute platforms.
• Comprehensive "Quality of Flight" metrics: success rate, energy per mission, distance traveled, flight time.
• Support for Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO).

Maintenance & Community

Contributors hail from Harvard, UT Austin, Google, and TU Delft. Maintainers Srivatsan Krishnan and Bardienus Duisterhof are contactable via email. The project encourages contributions to enhance usability and functionality.

Licensing & Compatibility

The README does not specify the software license. Tested on Windows 10; Linux support is planned.

Limitations & Caveats

Linux support is currently unavailable. The absence of a specified software license is a significant adoption risk for commercial or sensitive research use cases.