Modeling for robotics and pneumatic actuation using PDEs

Abstract

Successful robotic control depends on truthful and robust dynamic models. While systemidentification is generally employed to compute such models, there are challenges to applying system ID. In the context of dynamic object manipulation the difficulty arises from the non- linear nature of the contact phenomena. On the other hand, with pneumatically actuated robots, the challenge is devising a suitable model class that captures the intrinsic dynamics of the system. The primary goal of this thesis is to push the state of the art in those areas. The thesis makes the following contributions towards this goal: • A novel approach to the system identification problem that solves both the problems of estimation and system ID jointly. We show that these two problems are difficult to solve separately in the presence of discontinuous phenomena such as contacts. The problem is posed as a joint optimization across both trajectory and model parameters and solved via Newton’s method. • A novel way of modeling pneumatically actuated systems based on Computational Fluid Dynamics and implemented as a Partial Differential Equation solver that is augmented to handle full robotic systems. • Validation of that model and exploration of the advantages compared to previously employed models via the use of Reinforcement Learning in a simulated environment.3 • System Identification of PDE-based pneumatic model and demonstration of its advan- tages over the existing methods for modeling pneumatic systems. • Successful transfer of agents learned through Reinforcement Learning to real hardware.