Machine learning and policy optimization methods applied to compliant control.
Investigators: Prof. Toussaint, Prof. P. Xu
The main focus is on the application and empirical evaluation of control policy optimization methods for compliant soft tissue manipulation. In the case of compliant controllers, the impedance along any task space dimensions can be parameterized separately and depending on time. In general, this is a very high-dimensional policy space. This project will investigate methods to cope with such high-dimensional problems in specific applications, also exploiting the specific structure and assumptions that can be made in a particular application. Such applications include, for instance, slicing through or sliding along a soft tissue while compliantly keeping contact with a constraint, with objectives that relate to the interaction forces and deformations. The goals are learning methods that lead to high-performance manipulations with as few trial-and-errors as possible. Additionally, we will evaluate our policy optimization methods within the simulations developed in Focus Area A – and then validate the so optimized policies in the real domain.