Morphological Design and Control of a Bio-Inspired, Structurally Compliant Quadruped
From the viewpoint of evolution, vertebrates first accomplished locomotion via motion of the spine. Legs evolved later, to enhance mobility, but the spine remains central. Contrary to this, most robots have rigid torsos and rely primarily on movement of the legs for mobility. The force distributing properties of tensegrity structures presents a potential means of developing compliant spines for legged robots, with the goal of driving motion from the robots core. In addition, the increasing complexity of soft and hybrid-soft robots highlights the need for more efficient methods of minimizing machine learning solution spaces, and creative ways to ease the process of rapid prototyping.
In this thesis I present the process of morphological design for a tensegrity quadruped robot, the first to the author's knowledge, which I call MountainGoat, and its impact on controllable locomotion. All parts of the robot, including legs and spine, are compliant. Control is initially demonstrated with three variations of MountainGoat, focusing on actuation of the spine as central to the locomotion process. Following the general pattern of biological evolution, leg actuation is developed next. Additionally, to reduce the overall machine learning space, I present four different choices of muscle groups to actuate: three for a primarily spine-driven morphology of a tensegrity quadruped, called MountainGoat, and one for a primarily leg-driven variation of this quadruped, and compare the resulting differences in locomotion speed. Each iteration of design seeks to reduce the total number of active muscles, and consequently reduce the dimensionality of the problem for machine learning, while still producing effective locomotion. The reduction in active muscles seeks to simplify future rapid prototyping of the robot. For this portion of the thesis, two separate approaches to actuation, one primarily spine-driven and the other primarily leg-driven, are explored.
Locomotion for all models is aided by the use of central pattern generators, feedback control via a neural network, and a two-tiered machine learning approach involving the Monte Carlo method as well as genetic evolution for parameter optimization.