(or: where artificial intelligence meets robotics)
The popular image that comes to mind for many people when you think of AI is often a kind of humanoid robot, perhaps in a brilliant and elegant white body, playing failures or d ‘Other things. In many famous cases, science fiction stories depict artificial intelligence as embodied in a robot, or more broadly, in a synthetic body. This allows AI to interact with the physical world, to move, to carry out actions, etc. And that, too, means that the robot know how to move with dexterity as a human being.
In this series of messages, I will briefly talk about a project on which I work in the last month: How to use automatic learning / AI techniques to teach a robotic manipulator to movefollowing as closely as possible a trajectory of the desider.
Robotics and AI, often, are considered in university literature as two different subjects, and for a good reason. RoboticsIn the most classic sense, deals with the modeling and control of robotic manipulators or mobile robots. They are often necessary to perform specific and repetitive tasks, several times in an industrial framework, and do it with very high precision. The design of an architecture of control that reaches this task is not trivial, but these robots are rarely considered to be “intelligent”, able to do a preprogrammed task. On the other hand, Artificial intelligence is a branch of computer science, studying how to design algorithms that “Simulate intelligent behavior». There are many definitions of intelligence, but in this subject, it is often called the capacity to act according to the entries and the experience of the sensors and the possibility of dealing with the uncertainties and the events of insistence. It is therefore not surprising that, lately, these two scientific areas have found an intersection in what it is sometimes called Intelligent robotics.
A carefully designed automatic learning algorithm can learn to move in a Essay and errors fashion.
As I briefly mentioned in the introduction, robotic manipulators are often moved by control algorithms designed with precision, which are based on a Precise knowledge of structural and physical parameters manipulator. These include the angles between the joints, the lights of the links, the masses, the friction, etc. The more precise these parameters, the more precisely an control architecture can be designed. But what happens if these parameters are not known? This is where automatic learning (ML) is coming. ML allows computers to learn data directly and is intuitive to understand why it is useful in this case. Just like a human baby, a robot can Learn to move just by moving. A carefully designed automatic learning algorithm can learn to move in a Essay and errors mode, do many tests and analyze the effects of your actions. In our case, actions are couples applied to the joints, and the effects are how these joints move.
The dynamic model of a robot, that is to say that the mathematical model describing how the entries affect the movement of each articulation, can be very complex, especially when they treat robots with many degrees of freedom ( DOF). Links are often Dynamically coupled: The movement, speed and acceleration of one link affect the other in a complex way. There is Effects of inertia, centrifugal, coolis and gravitational effects, non -linear friction and so on. All these elements can be difficult to model mathematically. So the question is: Can the robot learn to move? In the next series of messages, I will show different approaches, techniques and algorithms that I developed with one of my colleagues. You can find all the code developed for this project in This GitHub repository.