(A) data-driven approach to modeling and control of mobile robots with nonlinear dynamics

Abstract

Robots navigating in diverse natural environments should be able to respond to irregular and dynamic environments. Interaction with the environment complicates robot motion modeling and control significantly. For instance, walking robots or aquatic robots have complex dynamic models because they are influenced by their contact with the ground or their interactions with water. Modeling and controlling these robots require a high level of expertise, and even this level of expertise has its limitations at times. In this dissertation, we present data-driven modeling and control methods for robots that exhibit nonlinear and complex behaviors as a new breakthrough. The proposed methods encompass wide aspects of data-driven approaches, including learning an accurate model for long-term prediction, obtaining high-quality data for generalized models, and augmenting data for improving modeling and control performance. To validate the proposed methods, data are collected using a cruise tour boat in field experiments and using a small robotic surface vehicle in a laboratory. Moreover, we demonstrate the utility of the proposed method by enhancing offline reinforcement learning performance in a variety of simulation environments. Although the experimental data are collected in aquatic environments, which have complex nonlinear properties, it is expected that the proposed algorithm will be applicable to a wide range of fields, including soft robots, as it is domain-independent.