M.Sc. Thesis
My research during my M.Sc. degree was focused on investigating how a robot's morphological characteristics, such as proportions and number of segments in its spine and legs, affect its performance in terms of speed, adaptability, and energy consumption. You can find a few of the projects I worked on on this page.
Co-evolution of Morphology and Controller for a Compliant Quadruped Robot
During my graduate studies, I worked under the supervision of Prof. Majid Nili Ahmadabadi on developing an evolutionary methodology to determine the morphological parameters of a compliant quadruped robot. The main focus of my research was on the importance of morphology in the system’s final behavior and performance characteristics. Meaning that, instead of solely depending on the control part to do the job, we try to design the morphology and hence change the natural dynamics of the robot; making it more compatible with the task it is supposed to do. Consequently, the desired task is a natural outcome of the interaction of the situated body and its environment, with a minimal required effort.
Deciding on the suitable values for the robot’s morphological parameters is a complex task. Robot designers require a scientific tool to observe the influence of these parameters on the output behavior to help them decide about their implementation. we introduced a simulator for quadruped robots using the Open Dynamics Engine (ODE) library to make the morphological study as simple as possible. It uniquely provides the opportunity to transfer morphological changes to the simulation instantly and to obtain performance characteristics such as transportation cost and the robot’s average speed.
A Human-aided Evolutionary Approach
Sometimes the outcome of an evolutionary algorithm is an amazing creature that performs the assigned task but does not look like what we have expected. These cases are interesting because they re-emphasize on the dilemma of deciding on an evaluation function for an evolutionary algorithm. In our problem, the dimensionality of the search space is high and the space is extremely scattered. Global optimization methods such as Genetic Algorithm will easily get stuck at a local optima. Here, we proposed a human-collaborating method, to narrow down the search space using the designer's knowledge about how the perfect locomotion behavior should look like. There are many arguments about the cost function of an evolutionary process in biology. What was the cost function of the natural evolution that has resulted in this spectrum of animals? How the symmetry in the animals' structures has been evolved in nature? Questions such as these are still not clearly answered. Using the Cost of Transportation (COT) as the only metric for the GA, we obtained creatures that were suitable, regarding their COT, but showed unnatural behavior. If we can not find a general formulation that encodes the desired natural behavior, how about using the concept that any human designer has in his mind in this regard? Therefore, Instead of trying to find a mathematical formulation for what a natural behavior should look like, we used human intuition to choose a set of acceptable individuals as an initial population for the GA.
An Analytical Viewpoint
The starting point of my graduate research was this basic idea: with a few realistic assumptions, we suppose that a legged robot can be modeled as an equivalent structure, composed of a mass manipulated by a black box. If we have a desired trajectory for the center of mass (CoM) of this body, how could we optimally design the black box to achieve the desired performance goals? The black box is, of course, the robot’s legs in the stance phase, including links and potential passive compliant elements. In the first phase of my studies, I tried to pursue an analytical approach. Considering leg joints to have passive parallel compliances, I presented a methodology to determine their optimal value to minimize the tracking error. The challenging issue in this work is to come up with a method for resolving the redundancy that rises in the structure.
For more information please refer to the fourth chapter of my thesis. Continuing this path led to two published articles, with my lab mates, which are available on my Google Scholar.
Tip-toe Walking Detection Using CPG Parameters from Skeleton Data
I collaborated in another research in which we used a proposed model of programmable Central Pattern Generators (CPGs) to extract the most relevant features for gait recognition, distinguishing tip-toe from normal walking patterns. This model was then used for screening movement abnormalities seen in children with Autism.
This work has been presented at UCAmI 2017 and won the best paper award. We are currently working on extending this method to other body motion characteristics, such as hand flapping, jerky motions, fluttering, and spinning, seen in Autistic children.