Distributed and Parallel Optimization of Captured Human Motions for the Generation of Robust and Stable Motions on Simulated Humanoid Robots

There are several areas of research in the field of motion learning for robots. RoboCanes, a research group belonging to the Department of Computer Science at University of Miami, uses the process of recording a human motion and optimizing the same to get a stable motion for a simulated robot in the framework for the agent code. The major downside of this approach is that this optimization process takes hours to return a good set of parameters which led to the need of a fast parallelizing approach. This thesis is about an extension of the existing framework to generate stable motions for simulated humanoid robots using a motion capture framework and then optimizing the motions efficiently in a distributed and parallel environment. This approach is based on a client server network of systems where the server system controls the optimization process and distributes the particles (candidate solutions of an optimization process) to multiple client systems, which are responsible for running simulations individually, using the parameters sent by the server and then evaluating the error and returning it to the server. The experiments are conducted on three different experimental setups with four motion files and three optimization algorithms. The performance of this technique is measured with varying number of clients, each of which show considerable speedup as compared to the serial process.