Physically consistent state estimation and system identification for contacts

Successful model based control relies heavily on proper system identification and accurate state estimation. We present a framework for solving these problems in the context of robotic control applications. We are particularly interested in robotic manipulation tasks, which are especially hard due to the non-linear nature of contact phenomena. We developed a solution that solves both the problems of estimation and system identification jointly. We show that these two problems are difficult to solve separately in the presence of discontinuous phenomena such as contacts. The problem is posed as a joint optimization across both trajectory and model parameters and solved via Newton´s method. We present several challenges we encountered while modeling contacts and performing state estimation and propose solutions within the MuJoCo physics engine. We present experimental results performed on our manipulation system consisting of 3-DOF Phantom Haptic Devices, turned into finger manipulators. Cross-validation between different datasets, as well as leave-one-out cross-validation show that our method is robust and is able to accurately explain sensory data.