AI-Enabled Control System Helps Autonomous Drones Stay on Target in Uncertain Environments

In challenging environments where unexpected forces like gusty winds can push autonomous drones off course, staying on target is critical. MIT researchers have developed a new machine learning-based adaptive control algorithm designed to minimize trajectory deviation in the face of such unpredictable disturbances.

Unlike traditional methods that require advance knowledge of potential disturbances, this innovative technique allows the drone’s AI model to learn from a small amount of observational data, specifically 15 minutes of flight time. The system autonomously selects the most suitable optimization algorithm to adapt to specific disturbances, enhancing tracking performance by tailoring its approach to the unique geometry of each situation.

The researchers employ meta-learning, which trains the control system to concurrently adapt to various types of disturbances. This combined approach results in a 50 percent reduction in trajectory tracking error compared to baseline methods in simulations. The system also demonstrated superior performance in dealing with new wind speeds not encountered during training.

According to Navid Azizan, the Esther and Harold E. Edgerton Assistant Professor in MIT’s Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), “The concurrent learning of these components is what gives our method its strength. By leveraging meta-learning, our controller can automatically make choices that will be best for quick adaptation.”

The control system uses a neural network model to approximate disturbances from data, eliminating the need for an a priori understanding of wind speeds. The algorithm also automatically selects the right mirror-descent function, optimizing performance without requiring manual user input.

To enhance flexibility, the researchers use meta-learning to train the controller by exposing it to a range of wind speed families. Sunbochen Tang, the lead author of the paper and a graduate student in the Department of Aeronautics and Astronautics, explains, “Our method can cope with different objectives because, using meta-learning, we can learn a shared representation through different scenarios efficiently from data.”

In simulations and real-world experiments, the new method significantly reduced trajectory tracking error compared to baseline approaches. The team is now testing the control system on real drones under varying wind conditions and plans to extend the method to handle disturbances from multiple sources, such as shifting parcel weight during flight.

Babak Hassibi, the Mose and Lillian S. Bohn Professor of Electrical Engineering and Computing and Mathematical Sciences at Caltech, commented, “Navid and his collaborators have developed breakthrough work that combines meta-learning with conventional adaptive control to learn nonlinear features and the suitable adaptation law from data…Their work can contribute significantly to the design of autonomous systems that need to operate in complex and uncertain environments.”

This research was supported by MathWorks, the MIT-IBM Watson AI Lab, the MIT-Amazon Science Hub, and the MIT-Google Program for Computing Innovation.