How can we enable safe aviation autonomy with machine-learning (ML) in the loop?

Future aviation systems, in the form of unmanned aircraft systems (UAS) and urban air mobility (UAM) services, will operate at increased levels of autonomy by extensively leveraging machine-learning-enabled components (LECs), in tasks as diverse as visual perception, intent prediction, and decision-making. Despite the promise of LECs to outperform their traditional, non-learning based counterparts, today’s LECs can be notoriously poor at generalizing beyond their training data, and largely lack appropriate methods for verification and validation. To enable wider and trusted adoption of LECs, this project identifies three technical challenges that must be overcome to achieve the vision of autonomous aviation systems with LECs in the loop.

Publicly-Available Simulator and Dataset

We have released a dataset, based on the X-Plane simulator, to stress test learning-based perception and control modules for aviation autonomy here with code on Github.