An autonomous drone carrying water to help turn off a forest fire in the Sierra Nevada could meet swirling winds from Santa Ana which threaten to push it. The rapid adaptation to these unknown disturbances, the influence presents a huge challenge for the drone flight control system.
To help such a drone stay on the target, MIT researchers have developed a new adaptive control algorithm based on automatic learning which could minimize its gap compared to its planned trajectory in the face of unpredictable forces such as burst winds.
Unlike standard approaches, the new technique does not require the person programming the autonomous drone to know anything in advance on the structure of these uncertain disturbances. Instead, the artificial intelligence model of the control system learns everything it needs to know that a small amount of observation data collected from 15 minutes of flight.
Above all, the technique automatically determines which optimization algorithm it should use to adapt to disturbances, which improves monitoring performance. He chooses the algorithm that best suits the geometry of specific disturbances with which this drone is confronted.
The researchers form their control system to do both things simultaneously using a technique called meta-learning, which teaches the system to adapt to different types of disturbances.
Together, these ingredients allow their adaptive control system to reach 50% lesson tracking error than basic methods in simulations and work better with new wind speeds that it has not seen during training.
In the future, this adaptive control system could help autonomous drones more efficiently deliver heavy plots despite strong winds or monitor areas subject to fire from a national park.
“The competitor Learning of these components is what gives Our Method its Strength. By Leveraging Meta-learning, our controller can automatically make choices that will be for Quick Adaptation,” Says Navid Azizan, Who is the Esther and Harold E. Edgerton Assistant Professor in the Mit Mechanical Engineering and the Institute for Data, Systems, and Society (IDSS), A Principal Investigator of the Laboratory for Information and Decision Systems (Courcles), and the main author of a paper on this control system.
Azizan is joined on the newspaper by the main author Sunbochen Tang, a student graduated from the department of aeronautics and astronautics, and Haoyuan Sun, a graduate student of the Department of Electrical and Computer Science. Research was recently presented at the Learning for Dynamics and Control conference.
Find the right algorithm
As a rule, a control system incorporates a function that models the drone and its environment, and includes certain existing information on the structure of potential disturbances. But in a real world fulfilled with uncertain conditions, it is often impossible to design this structure by hand in advance.
Many control systems use an adaptation method based on a popular optimization algorithm, called gradient descent, to estimate the unknown parts of the problem and determine how to keep the drone as close as possible to its target trajectory during flight. However, the gradient descent is only a algorithm in a larger family of available algorithms, called mirror descent.
“Mirror Descent is a general family of algorithms, and for a given problem, one of these algorithms can be more appropriate than the others. The name of the game is how to choose the particular algorithm that suits your problem. In our method, we automate this choice, ”says Azizan.
In their control system, researchers have replaced the function which contains a certain structure of potential disturbances by a neural network model which learns to approximate it from the data. In this way, they do not need to have an a priori structure of the wind speeds that this drone could meet in advance.
Their method also uses an algorithm to automatically select the mirror function of the good mirror while learning the neural network model from data, rather than assuming that a user already has the ideal function. The researchers give this algorithm a range of functions to choose, and it finds the one that best suits the problem at hand.
“The choice of a good distance generation function to build the right mirror-descent adaptation matters a lot to obtain the right algorithm to reduce the follow-up error,” adds Tang.
Learn to adapt
Although the wind accelerates that the drone can meet could change each time it takes off, the controller's neural network and the mirror function must remain the same so that they do not need to be recomposed each time.
To make their controller more flexible, researchers use meta-learning, teaching it to adapt by showing it a range of wind speed families during training.
“Our method can face different objectives because, using meta-learning, we can learn a shared representation through different scenarios from data,” explains Tang.
In the end, the user feeds the control system a target trajectory and continuously recalculates, in real time, how the drone should produce the thrust to keep it as close as possible to this trajectory while adapting to the uncertain disturbances it encounters.
In the simulations and experiences of the real world, the researchers have shown that their method led to an error in monitoring the trajectory clearly less than the basic approaches at each speed of the wind they tested.
“Even if wind disorders are much stronger than we have seen during training, our technique shows that it can always manage them successfully,” adds Azizan.
In addition, the margin by which their method surpassed the basic lines increased as the wind speeds intensified, showing that it can adapt to difficult environments.
The team now performs material experiences to test its control system on real drones with variable wind conditions and other disturbances.
They also want to extend their method so that they can manage the disturbances of several sources at the same time. For example, the modification of wind speeds could cause the weight of a package that the drone carries to change theft, especially when the drone carries useful slip loads.
They also want to explore continuous learning, so that the drone can adapt to new disturbances without having to recycle the data it has seen so far.
“Navid and his collaborators have developed revolutionary work that combines meta-learning with conventional adaptive control to learn nonlinear functionalities from data. The key to their approach is the use of mirror descent techniques that considerably exploit the design of autonomous systems that require a complex and uncertain environments, “explains Babak S. Bohn professor of electrical and computer engineering and mathematical sciences of Caltech, which was not involved in this work.
This research was supported, in part, by Mathworks, the Mit-ibm Watson Ai Lab, the Mit-Amazon Science Hub and the Mit-Google Program for Computer Science.
