Autonomous drones that help put out wildfires in Sierra Nevada may encounter swirling Santa Ana winds that threaten to push the course away. Inflights that quickly adapt to these unknown disturbances pose a major challenge for drone flight control systems.
To help such drones stay on target, MIT researchers have developed new machine learning-based adaptive control algorithms that can minimize deviations from the intended trajectory in the face of unpredictable forces like Gusty Winds.
Unlike standard approaches, the new approach does not require the autonomous drone programmer to know anything in advance about the structure of these uncertain obstacles. Instead, the artificial intelligence model of the control system learns everything you need to know from the small amount of observed data collected from the 15-minute flight time.
Importantly, this technique is to automatically determine the optimization algorithms that need to be used to improve tracking performance. Choose the best algorithm for the geometry of the specific disturbances this drone is facing.
Researchers train their control systems to do both things at the same time using a technique called Meta-Learning. This teaches the system how to adapt to different types of interference.
Together, these components allow adaptive control systems to have 50% fewer trajectory tracking errors than the baseline methods of the simulation, and to improve performance with new wind speeds that were not displayed during training.
In the future, this adaptive control system will help autonomous drones provide heavier compartments in greater efficiency and monitor fire-prone areas in national parks despite strong winds.
“Simultaneous learning of these components gives our methods to their strength. By leveraging meta-learning, controllers can automatically make the best choice for quick adaptation,” a senior author of this paper on control system.
Azizan is added to the paper by lead author Sunbochen Than, a graduate student in the Department of Aerospace and Space, and Haoyan Sang, a graduate student in the Department of Electrical Engineering and Computer Science. This study was recently presented at the Dynamics and Control Conference Study.
Find the right algorithm
Control systems typically incorporate functions that model the drone and its environment, and contain existing information about the structure of the potential fault. However, in a real world where uncertain conditions are met, it is often impossible to pre-draw this structure.
Many control systems use adaptation methods based on a common optimization algorithm known as gradient descent to estimate unknown parts of the problem and determine how to bring the drone as close as possible to the target trajectory in flight. However, gradient descent is just one algorithm in a larger family of algorithms to choose from. This is known as the mirror descent.
“Mirror Descend is a general family of algorithms, and for certain issues, one of these algorithms is more appropriate than the others. The name of the game is how to choose a specific algorithm that suits the problem. This method automates this selection,” says Azizan.
In control systems, researchers have replaced functions that include the structure of potential failures with neural network models that learn to approximate the data. In this way, there is no need to have an a priori structure of wind speeds that this drone may encounter beforehand.
Also, rather than assuming that the user has already selected it, it uses algorithms to automatically select the correct mirror decent function while learning the neural network model from the data. Researchers offer a variety of features to choose this algorithm and find the best feature at hand for the problem.
“Building the right mirror subsidence adaptation to select the right mirror generation function is extremely important in getting the right algorithm to reduce tracking errors,” adds Tang.
Learn to adapt
Wind speeds can change with each drone flight, but the controller's neural network and mirror functionality must remain the same, so there is no need to recalculate each time.
To make the controller more flexible, researchers use metal learning and teach them to adapt by showing different wind speed families during training.
“Metalearning allows you to efficiently learn shared representations from your data using a variety of scenarios, allowing you to address a variety of objectives,” explains Tang.
Ultimately, the user supplies the target trajectory to the control system and continuously recalculates in real time how the drone generates thrust and keeps it as close as possible to that trajectory as possible, whilst dealing with the uncertain disturbances it encounters.
In both simulations and real-world experiments, researchers have shown that the method has significantly fewer trajectory tracking errors than the baseline approach at all wind speeds tested.
“Even if the wind turbulence during training is much stronger than we saw, our techniques show that it can still handle them well,” adds Azizan.
Furthermore, as wind speeds are enhanced, margins that the method exceed the baseline indicate that it can be adapted to challenging environments.
The team is currently running hardware experiments to test the control system on real drones with various wind conditions and other obstacles.
We also want to extend our methods so that we can handle interference from multiple sources at once. For example, changing wind speeds can cause the weight of the parcels that the drone is moving in flight, especially if it is carrying a sloshing payload.
They also want to explore continuous learning, so drones can adapt to new disturbances without having to retrain data they have seen so far.
“Navid and his collaborators have developed a groundbreaking task of learning nonlinear features from data by combining meta-learning with traditional adaptive control. The key to their approach is the use of mirror-declining techniques that could not exploit the fundamental geometry of previous art. Professor Bourne of Electrical Engineering, Computing and Mathematical Sciences in California, who was not involved in this work.
This study was supported in part by the MIT-Google programs for Mathworks, Mit-IBM Watson AI Lab, Mit-Amazon Science Hub, and Mit-Google Innovation.
