Pastor Peter (Peter Pastor) from the University of Southern California in conjunction with colleagues develops software that allows the robot PR 2 learn new skills in motion – through the demonstration of these movements man. For example, the robot learned to pour a drink for only one demonstration. The developers have focused on the model, often repetitive movements, such as turning a door handle or prehension of the cup. Moreover, the robot can perform the same action, even if the objects are in different positions. To do this, developers create a library of motions that can be used by a robot.
In real life, the robot will never be faced with completely the same situation twice. Therefore, the lessons of the movement must be encoded so as to be adapted to the different initial situations and changing goals. Developers use Dynamic Movement Primitives (DMPs), which in one demonstration in a few seconds allows the user to effectively teach PR2 new movements. DMPS allows the robot to accomplish its task, even if the target is changed during the execution of movement.
However, as the researchers say, they need to solve three problems:
The first – “the problem of coordination”, as Mechanical scheme of the extremities (including joints) in the robot and human – are different. The second – “the problem of generalization” movements, it is impossible to teach a robot across a huge number of possible motions. The third – “the problem of reliability” movements – because of the constantly changing context and changing the movement’s goals.
It appears that one way to solve this problem – the fragmentation of movements to the isolation of “primitive”, simple movements and the development of algorithms that generate the complex motions of the chains of “primitive” movements, which are incorporated into the overall movements of the robot base. And, of course, the transfer of the management teams of motors a robot, given the fact that a large number of motors of the robot and its degrees of freedom requires the choice of the optimal trajectory, as well as serious coordination of all elements of the management and the robot mechanical system.
Note that the developers also believe that they have developed models can be compared with the behavior of biological objects (models) to find out “primitive motion in biological systems, as well as to understand how such” primitive movement represented in the brain of biological objects.