|M.Sc Student||Shirizly Alon|
|Subject||Comparison of CPG Control Methods for Walking Biped Robots|
|Department||Department of Mechanical Engineering||Supervisor||Professor Miriam Zacksenhouse|
Dynamic walkers, which rely on the natural dynamics of the robot, enjoy high efficiency and speed, by relinquishing the requirement of static or quasi-static stability at each point in the trajectory. Central Pattern Generators (CPGs) are prominent biologically inspired mechanisms to generate periodic movements.
In biology, CPGs have been shown to have two main functions: setting the rhythm of the movement and coordinating the pattern of activity of different joints. In 1-level CPGs, these two functions are performed by a single system that is generally based on a network of coupled oscillators, which receive continuous feedback signals and whose outputs are used directly as control signals for the motors. In 2-level CPGs, these two tasks are performed by two distinct subsystems, one to generate the rhythm, and another to generate the pattern of activity with that rhythm. 1-level and 2-level CPGs have been shown in the literature to generate robust and efficient robotic gaits, but their relative merits have not been compared. In this thesis I present a thorough comparison of the learning process and performance of 1-level and 2-level CPG.
A controller is designed to represent each of the two control classes: a 1-level CPG consisting of 3 coupled Matsuoka Oscillators (MOs), and a 2-level CPG using a resetting phase generator and rectangular pulse generators. The controllers generate control signals (torque to actuated joints) for a walking compass biped. The controllers' parameters are tuned using Genetic Algorithm (GA), optimizing 3 objectives: (1) Velocity over flat terrain, (2) Robustness over rough (stochastically generated) terrain, and (3) Robustness over increasing slopes. The GAs are repeated to generate statistically significant results, and these results are compared to show clear advantage to the 2-level controller in all 3 objectives (p-value below 0.05).
As mentioned, a common model for a CPG is based on a network of Matsuoka Oscillators: a simple network of self and mutual inhibiting neurons, each modeled as a second order non-linear system. This model demonstrates a rich variety of behaviors, and analytic conditions for generating periodic cycles have been found. Here I analyze limit-cycle stability using numerical and analytical tools, and investigate phenomena of period-adding and symmetry breaking bifurcations.