Data-Augmented Motion Planning and Predictive Control for Robotic Jumping on Highly Dynamical Terrains

Abstract

Model-based controllers have shown strong performance in a wide range of robotic systems, from relatively simple platforms such as drones and unicycle cars to more complex systems such as quadrupedal and bipedal robots. These controllers also provide a principled framework for synthe- sizing diverse behaviors in robotic systems, including jumping in hopping robots, trotting, pacing, and bounding in quadrupeds, and walking or running in humanoid robots. At the core of model-based control lies the use of dynamical models to predict future robot behavior based on current states and control inputs. However, discrepancies between the idealized models used for control and the real-world physical systems are often a major source of control failure. These mismatches can stem from various sources, most commonly from inaccurately modeled dynamics, such as rigid-body assumptions that ignore structural deformation and unmodeled dynamics, including compliance in robot-environment interactions (REI) and oversimplified representations of internal component behavior. To address these challenges, this thesis proposes a Data-Augmented Control (DAC) framework that leverages input-output data to model complex or qualitatively hard to define dynamics and integrates them into both motion planning and predictive control. By directly incorporating fitted interaction dynamics into the control pipeline, the DAC framework enables more dynamic, and efficient locomotion behaviors, particularly in scenarios involving compliant or unstructured environments.