Learning to Fly (Behavioural Cloning)

For several years, the AI Laboratory has had a research project aimed at using machine learning to acquire control skills based on examples provided by human trainers. One of the test domains has been in learning to fly an aircraft in a flight simulator. In these experiments, human pilots fly a given flight plan. Each time the pilot performs a control action, the action, along with the state of the simulation is recorded. Flight logs are used as input to learning programs that output control rules that can be put into the autopilot of the aircraft.

In cooperation with the School of Mechanical and Manufacturing Engineering, we have access to a full cockpit simulator and real pilots for collecting flight data.

This project provides scope for work in advanced simulation graphics as well as machine learning.