Model learning for trajectory tracking of robot manipulators

Abstract Model based controllers have drastically improved robot performance, increasing task accuracy while reducing control effort. Nevertheless, all this was realized with a very strong assumption: the exact knowledge of the physical properties of both the robot and the environment that surrounds it. This assertion is often misleading: in fact modern robots are modeled in a very approximate way and, more important, the environment is almost never static and completely known. Also for systems very simple, such as robot manipulators, these assumptions are still too strong and must be relaxed. Many methods were developed which, exploiting previous experiences, are able to refine the nominal model: from classic identification techniques to more modern machine learning based approaches. Indeed, the topic of this thesis is the investigation of these data driven techniques in the context of robot control for trajectory tracking. In the first two chapters, preliminary knowledge is provided on both model based controllers, used in robotics to assure precise trajectory tracking, and model learning techniques. In the following three chapters, are presented the novelties introduced by the author in this context with respect to the state of the art: three works with the same premise (an inaccurate system modeling), an identical goal (accurate trajectory tracking control) but with small differences according to the specific platform of application (fully actuated, underactuated, redundant robots). In all the considered architectures, an online learning scheme has been introduced to correct the nominal feedback linearization control law. Indeed, the method has been primarily introduced in the literature to cope with fully actuated systems, showing its efficacy in the accurate tracking of joint space trajectories also with an inaccurate dynamic model. The main novelty of the technique was the use of only kinematics information, instead of torque measurements (in general very noisy), to online retrieve and compensate the dynamic mismatches. After that the method has been extended to underactuated robots. This new architecture was composed by an online learning correction of the controller, acting on the actuated part of the system (the nominal partial feedback linearization), and an offline planning phase, required to realize a dynamically feasible trajectory also for the zero dynamics of the system. The scheme was iterative: after each trial, according to the collected information, both the phases were improved and then repeated until the task achievement. Also in this case the method showed its capability, both in numerical simulations and on real experiments on a robotics platform. Eventually the method has been applied to redundant systems: differently from before, in this context the task consisted in the accurate tracking of a Cartesian end effector trajectory. In principle very similar to the fully actuated case, the presence of redundancy slowed down drastically the learning machinery convergence, worsening the performance. In order to cope with this, a redundancy resolution was proposed that, exploiting an approximation of the learning algorithm (Gaussian process regression), allowed to locally maximize the information and so select the most convenient self motion for the system; moreover, all of this was realized with just the resolution of a quadratic programming problem. Also in this case the method showed its performance, realizing an accurate online tracking while reducing both the control effort and the joints velocity, obtaining so a natural behaviour. The thesis concludes with summary considerations on the proposed approach and with possible future directions of research.