Please use this identifier to cite or link to this item:
Authors: Ruppert, Felix
Title: Engineered bioinspired natural dynamics and their synergy with control and learning in legged robots
Issue Date: 2022 Dissertation
Abstract: The performance of legged locomotion relies on the successful mitigation of unstructured, rough terrain in the presence of sparse information and neurosensory delays. Bioinspired walking systems benefit from carefully engineered passive compliant behavior that models the inherent elastic behavior of muscle-tendon structures in animals. To leverage the passive behavior that provides energy efficiency, passive stability as well as simplified control and learning tasks to the system, locomotion control and learning algorithms have to be designed and coordinated with the natural system dynamics in mind to achieve similar locomotion behavior we see in animals. The major contribution of this thesis is the synergy of a bio-inspired leg design with biarticular muscle-tendon structures, a wearable force and pressure sensor design for closed-loop control in legged locomotion, a biologically inspired closed-loop central pattern generator with reflex-like feedback and a learning approach that enables the locomotion controller to leverage the carefully engineered natural dynamics of the robot to learn convincing locomotion skills and increase energy efficiency. The first contribution is a biologically inspired leg design focusing on the biarticular lower leg muscle-tendon structure in vertebrate animals. The biarticular elasticity provides two-dimensional passive impedance to the leg and allows the storage of energy orthogonal to the leg axis direction. The leg blueprint is characterized in its capability to store and release elastic energy in the biarticular structure. The stored energy can be recuperated back into the system and increases the energy efficiency of the leg. This leg design achieves the lowest relative cost of transport documented for all dynamically hopping and running robots. The second contribution introduces the concept of training wheels, temporary mechanical modifications to the system dynamics that shape the learning reward landscape and simplify learning locomotion directly in hardware. Through deliberate changes to the system dynamics, in this case, reduced gravity, the reward landscape can be shaped to simplify the learning process. Learning with this training wheel is safer due to smoother reward landscapes with shallower gradients. Also, the initial guess for initiating the machine learning algorithm is simplified, because the salient gradient set of viable reward signals is bigger. During the learning process, the training wheel influence can be gradually reduced and the learning algorithm converges to the solution of the initial learning problem without training wheels. The third contribution presents a rugged, lightweight force and pressure sensor for feedback information and biomechanical analysis. The sensor can be mounted on a robotic foot and provides continuous force and pressure feedback during locomotion in unstructured and soft terrain. The sensor is based on a pressure sensor, encapsulated in a polyurethane dome with an air cavity around the pressure sensor. External forces deform the sensor dome and the rising pressure in the air cavity is measured by the pressure sensor. Based on the dome geometry and material, the sensor range can be adjusted for different load cases. The sensor can be used in arrays to measure pressure distributions as well as a wearable force sensor in wet or granular media where classical force plates can not be utilized. The final contribution synergizes the individual contributions into a neuroinspired learning approach that matches a bioinspired closed-loop central pattern generator with reflex-like neuroelastic feedback to the natural dynamics of a quadruped robot with biarticular legs. Through sparse contact feedback from the foot sensor, the bioinspired central pattern generator can neuroelastically mitigate short-term perturbations to adapt the robot to its environment. Because the robot dynamics and the control task dynamics initially do not match, the controller uses the neuroelastic feedback to minimize the discrepancy between commanded and measured robot behavior. The amount of feedback activity during level walking can be used as a proxy to estimate the amount of dynamics mismatching. By minimizing the amount of required neuroelastic feedback the robot learns to neuroplastically match its control task dynamics to its natural dynamics through Bayesian optimization. Through the synergy of mechanics and control the biomechatronic system benefits from both the individual functionality of its components as well as their interplay during locomotion. The designed natural dynamics provide advantageous passive behavior to the robot and the bioinspired controller learns to leverage the natural dynamics to achieve convincing locomotion.
Appears in Collections:02 Fakultät Bau- und Umweltingenieurwissenschaften

Files in This Item:
File Description SizeFormat 
main.pdf12,05 MBAdobe PDFView/Open

Items in OPUS are protected by copyright, with all rights reserved, unless otherwise indicated.