Machine Learning in Sequential Composition of Control
|Project members:||dr. E. Najafi, MSc (Esmaeil), dr. G.A.D. Lopes (Gabriel), prof.dr. R. Babuška (Robert)|
|Keywords:||Robotics and mechatronics, Intelligent control, Hybrid and nonlinear systems|
Designing motion controllers for ground robots that need to traverse unstructured terrains is a challenging problem. This is due to the fact that such controllers should able to cope with modeled and unmodeled situations. For example, a mobile robot should be able to execute different tasks such as walking, jumping, running, going up stairs, in different types of terrain. This class of problems cannot be solved by linear controllers and nonlinear controllers must be used. In this framework each task requires a specialized nonlinear controller. A supervisory mechanism is then needed to compose the specialized controllers. This is called sequential composition of controllers and such an approach requires a supervisor whose task is to switch between the controllers in appropriate way. In the formulation of sequential composition of controllers, when the system reaches its goal which has been designed to lie inside the domain of attraction, a transition event occurs between discrete supervisor states.
This project explores automatic synthesis of composable controllers and of the supervisor. Some challenges are: 1) Describing the domain of attraction for each controller, particularly for nonlinear controllers; 2) Describing the goal set for each controller; 3) Solving the problem of when the desired goal does not lie within the domain of attraction of any controller; 4) Solving the problem of when there is no intersections between the domains of attraction of the controllers; 5) Employ model based and learning approaches to enlarge the domains of attraction.
Three case study scenarios will be considered within this project which are: a) An underactuated 1-DOF nonlinear system; b) Mode transition in an autonomous six-legged mobile robot; c) Negotiation of complex terrains on the same robot.