Model Reference Adaptive Control for Robotic Manipulation with Kalman Active Observers

Nowadays, the presence of robotics in human-oriented applications demands control paradigms to face partly known, unstructured and time-varying environments. Contact tasks and compliant motion strategies cannot be neglected in this scenario, enabling safe, rewarding and pleasant interactions. Force control, in its various forms, requires special attention, since the task constraints can change abruptly (e.g. free-space to contact transitions) entailing wide variations in system parameters. Modeling contact/non-contact states and designing appropriate controllers is yet an open problem, even though several control solutions have been proposed along the years. Two main directions can be followed, depending on the presence of force sensors. The perception of forces allows explicit force control (e.g. hybrid position/force control) to manage the interaction imposed by the environment, which is in general more accurate than implicit force control schemes (e.g. impedance control) that do not require force sensing. A major problem of force control design is the robustness to disturbances present in the robotic setup. In this context, disturbances include not only system and measurement noises but also parameter mismatches, nonlinear effects, discretization errors, couplings and so on. If the robot is interacting with unknown objects, "rigid" model based approaches are seldom efficient and the quality of interaction can be seriously deteriorated. Model reference adaptive control (MRAC) schemes can have an important role, imposing a desired closed loop behavior to the real plant in spite of modeling errors. This chapter introduces the Active Observer (AOB) algorithm for robotic manipulation. The AOB reformulates the classical Kalman filter (CKF) to accomplish MRAC based on: 1) A desired closed loop system. 2) An extra equation to estimate an equivalent disturbance referred to the system input. An active state is introduced to compensate unmodeled terms, providing compensation actions. 3) Stochastic design of the Kalman matrices. In the AOB, MRAC is tuned by stochastic parameters and not by control parameters, which is not the approach of classical MRAC techniques. Robotic experiments will be presented, highlighting merits of the approach. The chapter is organized as follows: After the related work described in Section 3, the AOB concept is analyzed in Sections 4, 5 and 6, where the general algorithm and main design issues are addressed. Section 7 describes robotic