Abstract

This dissertation addresses the challenge of enabling soft robots to perform contact-rich manipulation tasks by overcoming limitations in hardware, modeling, control, and learning. First, we present hardware advancements with PneuDrive, a pneumatic control system for actuating soft robots. We also introduce Baloo, a hybrid soft-rigid robotic torso which achieves a single-arm payload capacity of 19.3 kg and robust whole-body grasping capabilities. We also develop a full-body tactile sensing system enabling piezoresistive sensor arrays to cover large areas (1 square meter) with over 8000 distributed sensing points. Second, we focus on developing tractable dynamic models based on the Recursive Newton-Euler algorithm for soft continuum actuators. We also present a lumped-parameter, high-speed MuJoCo simulation of Baloo, which runs up to 350x real time, laying the foundation for data-intensive learning algorithms. Third, we present several methods to achieve improved control performance on soft robots despite their dynamic and kinematic uncertainties. We develop a data-efficient hybrid modeling approach that combines physics-based models with a deep neural network for real-time error compensation. We also adapt a neural-network based adaptive controller with theoretical stability guarantees for large-scale soft robot arms and demonstrate real-time compensation under uncertainty and perturbation. Finally, we demonstrate the application of reinforcement learning for whole-body manipulation skills. We achieve an 88% success rate with a zero-shot, sim-to-real transfer of a learned policy to hardware. We show that guiding the learning process with a motion primitive accelerates and stabilizes training for this long horizon, multi-step task. Despite introducing additional failure modes, this policy exhibits beneficial learned reactive behaviors like re-grasping and perturbation recovery.

Degree

PhD

College and Department

Ira A. Fulton College of Engineering; Mechanical Engineering

Rights

https://lib.byu.edu/about/copyright/