CRII: RI: Semiparametric Approaches to Learning Robot Dynamics

点击翻译按钮获取中文摘要

Robotics is revolutionizing quality of life in a wide range of domains from healthcare to automobile safety. Parametric modeling techniques are fundamental to robotic prediction and control in these domains, employed with great success when a robot's interaction with an environment can be precisely characterized by Newtonian physics. As complex robotic technology moves into natural and human environments, it is becoming more difficult to robustly characterize these interactions a priori with parametric models. As a result, machine learning is an increasingly important tool: models of complicated and noisy dynamics can be directly learned from a robot's interaction with its environment. In particular, nonparametric learning has shown exceptional promise, often outperforming parameterized, physics-based models when applied to difficult modeling problems. However, nonparametric approaches also have practical drawbacks: they do not incorporate prior knowledge such as physics-based insights and constraints, they are data-intensive, and they often incur significant computational cost. This is an exploratory investigation of how nonparametric statistical models can be better integrated with parametric physics-based models for robot prediction and control. The focus of this project is on developing new semiparametric models that elegantly compose parametric and nonparametric components for accurate, robust modeling.