Bayesian Modeling for Optimization and Control in Robotics

Abstract

Robotics has the potential to be one of the most revolutionary technologies in human history. The impact of cheap and potentially limitless manpower could have a profound influence on our everyday life and overall onto our society. As envisioned by Iain M. Banks, Asimov and many other science fictions writers, the effects of robotics on our society might lead to the disappearance of physical labor and a generalized increase of the quality of life. However, the large-scale deployment of robots in our society is still far from reality, except perhaps in a few niche markets such as manufacturing. One reason for this limited deployment of robots is that, despite the tremendous advances in the capabilities of the robotic hardware, a similar advance on the control software is still lacking. The use of robots in our everyday life is still hindered by the necessary complexity to manually design and tune the controllers used to execute tasks. As a result, the deployment of robots often requires lengthy and extensive validations based on human expert knowledge, which limit their adaptation capabilities and their widespread diffusion. In the future, in order to truly achieve an ubiquitous robotization of our society, it is necessary to reduce the complexity of deploying new robots in new environments and tasks. The goal of this dissertation is to provide automatic tools based on Machine Learning techniques to simplify and streamline the design of controllers for new tasks. In particular, we here argue that Bayesian modeling is an important tool for automatically learning models from raw data and properly capture the uncertainty of the such models. Automatically learning models however requires the definition of appropriate features used as input for the model. Hence, we present an approach that extend traditional Gaussian process models by jointly learning an appropriate feature representation and the subsequent model. By doing so, we can strongly guide the features representation to be useful for the subsequent prediction task. A first robotics application where the use of Bayesian modeling is beneficial is the accurate learning of complex dynamics models. For highly non-linear robotic systems, such as in presence of contacts, the use of analytical system identification techniques can be challenging and time-consuming, or even intractable. We introduce a new approach for learning inverse dynamics models exploiting artificial tactile sensors. This approach allows to recognize and compensate for the presence of unknown contacts, without requiring a spatial calibration of the tactile sensors. We demonstrate on the humanoid robot iCub that our approach outperforms state-of-the-art analytical models, and when employed in control tasks significantly improves the tracking accuracy. A second robotics application of Bayesian modeling is automatic black-box optimization of the parameters of a controller. When the dynamics of a system cannot be modeled (either out of complexity or due to the lack of a full state representation), it is still possible to solve a task by adapting an existing controller. The approach used in this thesis is Bayesian optimization, which allows to automatically optimize the parameters of the controller for a specific task. We evaluate and compare the performance of Bayesian optimization on a gait optimization task on the dynamic bipedal walker Fox. Our experiments highlight the benefit of this approach by reducing the parameters tuning time from weeks to a single day. In many robotic application, it is however not possible to always define a single straightforward desired objective. More often, multiple conflicting objectives are desirable at the same time, and thus the designer needs to take a decision about the desired trade-off between such objectives (e.g., velocity vs. energy consumption). One framework that is useful to assist in this decision making is the multi-objective optimization framework, and in particular