1005 Alcyon Dr Bellmawr NJ 08031
+1 484-446-2566
support@bmcpros.com
An autonomous drone carries water to help put out wildfires in the Sierra Nevada, which could encounter a spinning Santa Ana wind threatening to push it away. Quickly adapting to these unknown interference functions poses a huge challenge to the flight control system of the drone.
To help such drones stay on target, MIT researchers have developed a new, machine-based adaptive control algorithm that can deviate it from its expected trajectory, which is facing unpredictable forces such as storm winds.
Unlike standard methods, new technologies do not require people who program autonomous drones to understand the structure of these uncertain interferences in advance. Instead, the AI model of the control system learns everything that needs to be known from a small amount of observation data collected from a 15-minute flight time.
Importantly, the technology automatically determines which optimization algorithm should be used to adapt to interference, thereby improving tracking performance. It chose the algorithm that best suits the specific interference geometry faced by the drone.
Researchers train their control system using a technique called meta-learning to accomplish both things at the same time, which teaches the system how to adapt to different types of perturbations.
To sum up, these components reduce the trajectory tracking error of its adaptive control system by 50% compared to the baseline method in simulation and better new wind speeds that were not seen during training.
In the future, this adaptive control system could help autonomous drones deliver heavy packages more efficiently despite strong winds or monitoring fire areas in national parks.
“By leveraging meta-learning, our controllers can automatically make the choices that are best suited for fast adaptation,” said Navid Azizan, and senior author of the control system paper.
Azizan is joined by Sunbochen Tang, graduate student in the Department of Aeronautics and Actronautics, and Haoyuan Sun, graduate student in the Department of Electrical Engineering and Computer Science. The study was recently presented at the Dynamics and Control conference.
Find the correct algorithm
Typically, the control system combines a function that models the drone and its environment and contains some existing information about the potential interference structure. However, in a real world full of uncertain conditions, it is usually impossible to design such structures in advance.
Many control systems use adaptation methods based on popular optimization algorithms, called gradient descent, to estimate unknown parts of the problem and determine how to bring the drone as close to its target trajectory as possible during flight. However, gradient descent is just one of the larger family of algorithms to choose from, called mirror descent.
“Mirror drop is a general family of algorithms, and for any given problem one of these algorithms may be more appropriate than the others. The name of the game is how to choose a specific algorithm that suits your problem. In our approach, we can automate this choice,” Azizan said.
In their control system, the researchers replaced functions containing some potentially disturbing structures with neural network models that learned to approximate them from the data. In this way, they do not need the previous wind speed structure, and the drone may encounter it in advance.
Their method also uses an algorithm to automatically select the correct mirrored prose function while learning neural network models from the data, rather than assuming that the user has selected the ideal function. The researchers provide the algorithm with a range of features to choose from and have found the most suitable features for the current problem.
“Choose a good distance power generation function to build the correct mirror detection adaptability is important for obtaining the correct algorithm to reduce tracking errors,” Tang added.
Learning to adapt
Although the wind speeds that the drone may encounter may change on each flight, the controller’s neural network and mirroring capabilities should remain the same so that they are not required to be recalculated every time.
To make their controller more flexible, the researchers used meta-learning to teach it to adapt to a range of wind speed families during training.
“Our approach can cope with different goals because using meta-learning, we can effectively learn shared representations from different schemes through data,” Tang explained.
Finally, the user provides the control system with a target trajectory, and it can continuously re-estimate in real time how the drone should generate thrust to get it as close to the trajectory as possible while accommodating uncertain interference.
In simulation and real-world experiments, the researchers showed that their method resulted in significantly less trajectory tracking errors at each wind speed they tested compared to the baseline method.
“Even if the winds are much stronger than we see in training, our technology can still handle them successfully,” Azizan added.
Furthermore, as wind speeds increase, its approach is more advantageous than baseline margins, suggesting that it can adapt to challenging environments.
The team is now conducting hardware experiments to control the system of real drones in different situations of risk and other interference.
They also want to extend their approach so that they can deal with interference from multiple sources immediately. For example, wind speeds may cause the weight of the package carried by the drone to move in flight, especially when the drone carries the payload.
They also want to explore continuous learning so that drones can adapt to new interference without retraining in the data seen so far.
“Navid and his collaborators have developed groundbreaking work that combines meta-learning with conventional adaptive control to learn nonlinear features from data. The key to its approach is to use mirrored descent techniques that exploit the basic geometry of the problem, work that cannot be done in previous artistic ways, and their work cannot create important work for uncertain environments. Bohn is not involved in this work, professor of electrical engineering, computational and mathematical sciences at Caltech.
This study was supported in part by Mathworks, MIT-IBM Watson AI Labs, MIT-Amazon Science Hub, and MIT-Google Computing Innovation Programs.