An independent drone transporting water to assist in extinguishing a wildfire in the Sierra Nevada may encounter swirling Santa Ana winds that endanger its course. Rapidly adjusting to these unforeseen disturbances during flight poses a significant challenge for the drone’s flight control system.
To aid such a drone in maintaining its target, MIT researchers have created a novel machine-learning-based adaptive control algorithm designed to minimize deviations from its intended path amidst unpredictable forces like turbulent winds.
Unlike traditional methods, this innovative technique doesn’t require the operator programming the autonomous drone to have any prior knowledge about the characteristics of these unknown disturbances. Rather, the control system’s artificial intelligence model acquires all necessary information from a small set of observational data collected over just 15 minutes of flight time.
Crucially, the method autonomously identifies which optimization algorithm to employ in adapting to the disturbances, thereby enhancing tracking performance. It selects the algorithm that best fits the particular geometry of the specific disturbances the drone encounters.
The researchers simultaneously train their control system to perform both tasks using a technique known as meta-learning, which instructs the system on how to adjust to various types of disturbances.
Collectively, these elements empower their adaptive control system to accomplish 50 percent less trajectory tracking error compared to baseline methods in simulations while also performing better with new wind speeds that were not encountered during training.
In the future, this adaptive control system could enable autonomous drones to deliver heavy packages more effectively even in strong winds or supervise fire-prone areas within a national park.
“The simultaneous learning of these components is what lends our method its power. By utilizing meta-learning, our controller can automatically make selections that are best suited for swift adaptation,” states Navid Azizan, the Esther and Harold E. Edgerton Assistant Professor in the MIT Department of Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), and a principal investigator at the Laboratory for Information and Decision Systems (LIDS), as well as the senior author of a paper detailing this control system.
Azizan is joined in the paper by lead author Sunbochen Tang, a graduate student in the Department of Aeronautics and Astronautics, along with Haoyuan Sun, a graduate student in the Department of Electrical Engineering and Computer Science. Their research was recently showcased at the Learning for Dynamics and Control Conference.
Selecting the appropriate algorithm
Typically, a control system incorporates a function that models the drone and its surroundings, which includes some pre-existing information about the structure of potential disturbances. However, in a real world filled with unpredictable conditions, it can often be impossible to manually design this structure in advance.
Numerous control systems utilize an adaptation method grounded in a well-known optimization algorithm called gradient descent, allowing for the estimation of unknown aspects of the problem and figuring out how to keep the drone as near as possible to its target trajectory during flight. Nevertheless, gradient descent represents only one algorithm in a broader family termed mirror descent.
“Mirror descent encompasses a wide array of algorithms, and for any specific problem, one of these algorithms may prove to be more appropriate than others. The crux of the matter is how to select the precise algorithm that is optimal for your issue. In our method, we automate this selection process,” Azizan states.
Within their control system, the researchers substituted the function that harbors some structure of potential disturbances with a neural network model that learns to approximate them based on data. In this manner, they do not require an a priori structure of potential wind speeds that the drone could face.
Their approach also incorporates an algorithm designed to automatically choose the correct mirror-descent function while simultaneously learning the neural network model from data, rather than presuming a user has already selected the ideal function. The researchers provide this algorithm with a range of functions to select from, which it evaluates to find the one that aligns best with the current problem.
“Choosing an effective distance-generating function to construct the appropriate mirror-descent adaptation is crucial in obtaining the right algorithm to minimize tracking error,” Tang remarks.
Learning to adjust
While the wind speeds the drone may experience could vary with each flight, the controller’s neural network and mirror function should remain constant, eliminating the need for recalculation every time.
To enhance the flexibility of their controller, the researchers employ meta-learning, training it to adapt by presenting a range of wind speed families throughout the training phase.
“Our technique can manage various objectives because, through meta-learning, we can efficiently derive a shared representation from diverse scenarios using data,” Tang elaborates.
Ultimately, the user inputs a target trajectory into the control system, which continually recalibrates, in real-time, how the drone must generate thrust to remain as close as possible to that trajectory while accommodating the unpredictable disturbances encountered.
In both simulations and practical experiments, the researchers demonstrated that their method resulted in significantly lower trajectory tracking error compared to baseline approaches across all wind speeds tested.
“Even when wind disturbances are considerably stronger than what we experienced during training, our method proves capable of managing them effectively,” Azizan adds.
Moreover, the margin by which their technique surpassed the baselines increased as wind speeds intensified, indicating its adaptability to challenging environments.
The team is currently conducting hardware tests to evaluate their control system on actual drones under varying wind conditions and other disturbances.
Additionally, they aim to expand their method to handle disturbances from multiple sources simultaneously. For example, shifting wind speeds could cause the weight of a parcel carried by the drone to shift during flight, particularly when transporting sloshing payloads.
They also aspire to explore continual learning, facilitating the drone’s adaptation to new disturbances without retraining on previously seen data.
“Navid and his colleagues have achieved groundbreaking work that merges meta-learning with conventional adaptive control to learn nonlinear features from data. A pivotal aspect of their approach is the utilization of mirror descent techniques that take advantage of the underlying geometry of the issue in ways previous methods could not. Their research holds significant potential for the design of autonomous systems required to operate in intricate and uncertain environments,” states Babak Hassibi, the Mose and Lillian S. Bohn Professor of Electrical Engineering and Computing and Mathematical Sciences at Caltech, who was not involved with this study.
This research received partial support from MathWorks, the MIT-IBM Watson AI Lab, the MIT-Amazon Science Hub, and the MIT-Google Program for Computing Innovation.