An autonomous drone transporting water to help the peep in Sierra Nevada's fire may encounter Santa Ana's winding, which threatens her pushing. Quickly adapting to these unknown disorders creates a huge challenge for the drone flight control system.
To help such a drone, MIT researchers have developed a new algorithm for adaptation control based on machine learning, which could minimize its deviation from the intended trajectory in the face of unpredictable forces, such as a gusty wind.
Unlike standard approaches, the new technique does not require that the person programming the autonomous drone previously knows about the structure of these uncertain interference. Instead, the artificial intelligence model of the control system learns everything he must know from a small amount of observation data collected from 15 minutes of flight.
Importantly, the technique automatically determines which optimization algorithm should use to adapt to interference, which improves tracking efficiency. He chooses an algorithm that best suits the geometry of specific interference in front of which the drone stands.
Scientists train their control system to do both things at the same time using a technique called meta-learning, which teaches the system to adapt to different types of interference.
To sum up, these ingredients enable their adaptive control system to achieve a 50 percent less trajectory tracking error than output methods in simulations and achieving better wind speeds that he did not see during training.
In the future, this adaptive control system can help in autonomous drones more efficiently supply heavy plots despite strong winds or monitor the field areas of the national park.
“The coexisting learning of these components gives our method of its strength. By using the metal, our controller can automatically make choices that will be best for quick adaptation,” says Navid Azizan, who is Ester and Harold E. Edgerton, assistant to the MIT engineering in the field of mechanical engineering and the Institute of Data, Systems and Society Elder author A paper in this control system.
Azizan is joined by the main author of Sunbochen Tang, a graduate of the Aeronautics and Astronautics Department and Haoyuan Sun, a graduate of the Department of Electrical Engineering and Information Technology. The research has recently been presented at the Learning for Dynamics and Control conference.
Finding the right algorithm
Usually, the control system contains a function that models the drone and its environment, and contains some existing information on the structure of potential interference. But in the real world filled with uncertain conditions, it is often impossible to develop this structure in advance.
Many control systems use an adaptation method based on a popular optimization algorithm, known as the origin of the gradient, to estimate the unknown parts of the problem and determine how to keep the drone as close as possible to its target trajectory as possible during the flight. However, the descent of the gradient is only one algorithm in a larger family of available algorithms to choose from, known as a mirror fall.
“Descent Mirror is a general family of algorithms, and for each problem one of these algorithms can be more suitable than others. The name of the game is to choose a specific algorithm that is suitable for your problem. In our method we automate this choice,” says Azizan.
In their control system, scientists have replaced a function containing a certain structure of potential interference using a neural network model, which learns to bring them closer on the basis of data. In this way, they do not have to have the wind speed of wind speed, which this drone may encounter in advance.
Their method also uses an algorithm to automatically select the right mirror function when learning the neural network model based on data, instead of assuming that the user already has a chosen ideal function. Scientists give this algorithm a number of functions to choose from, and finds the one that suits the problem best.
“Choosing a good distance generation function to construct the right adaptation of the mirror has a lot of validity in obtaining an appropriate algorithm to reduce tracking error,” adds Tang.
Learning to adapt
While the wind of the speeds with which the drone encountered, it can change every time it escapes, the neural network and the function of the controller mirror should remain the same so that they do not need to be registered every time.
To make their controller more flexible, scientists use meta-tuition, teaching it to adapt, showing its series of wind speed families during training.
“Our method can cope with various goals, because using meta-tuition we can learn a shared representation through various scenarios based on data,” explains Tang.
Ultimately, the user powers the target trajectory control system and is constantly going well, how the drone should produce a string to keep it as close to this trajectory, while taking into account the uncertain disturbance he encounters.
In both simulations and experiments in the real world, scientists have shown that their method led to a much smaller trajectary tracking error than output approach at each wind speed tested.
“Even if wind disorders are much stronger than we have seen during training, our technique shows that it can still deal with them,” adds Azizan.
In addition, the margin, in which their method exceeded the base lines, grew with the severity of the wind speed, showing that it can adapt to difficult environments.
The team is now conducting hardware experiments to test their control system on real drones with different wind conditions and other disorders.
They also want to expand their method so that it can handle disturbances from many sources at the same time. For example, changing the wind speed can cause the weight of the plot that the drone transfers in the flight, especially when the drone transfers loads.
They also want to examine continuous learning so that the drone can adapt to new interference without the need to convince the data he has seen so far.
“Navid and his colleagues have developed groundbreaking works that combine meta-learning with conventional adaptive control in order to learn non-linear functions with data. The key to their approach is the use of mirror techniques that use the basic geometry of the problem in a way that could not be significantly. environment, BAF-MOSE, MOSE and LIS.
These studies have been partly supported by Mathworks, MIT-IBM Watson Ai Lab, Mit-Amazon Science Hub and Mit-Google program to calculate innovation.
