Optimizing Simulated Humanoid Robot Skills

摘要

Controlling a humanoid robot with a large number of joints and thus, degrees of freedom, presents a complex problem that requires knowledge in multiple fields, including biology, mechanics, physics, electronics and computer engineering. Soccer, a task that was created specifically for humans, requires not just precise and complex motor skills from an individual, but also strategy, planning and cooperation between the elements of the same team. The technologies involved in allowing a team of robots to play soccer span the main areas of artificial intelligence such as design principles of autonomous agents, multiagent collaboration, strategy acquisition, real-time reasoning and planning and intelligent robotics. The potential social and economic impact of significant developments in these areas coupled with the benefits of having a standard problem for researchers across the world to compete and collaborate, ultimately lead to the creation of the RoboCup initiative. This thesis aims to improve the performance of motor skills such as getting up, kicking the ball and walking developed by the FC Portugal3D team by applying to them an automated optimization process. The robot used in this work is the standard simulated RoboCup 3D robot, a simulated Nao. The advantages inherent to the use of a simulated robot, as opposed to a real robot, facilitate the development of this type of solution as well as the execution of subsequent experiments. Several different optimization algorithms, in particular Hill Climbing, Simulated Annealing, Tabu Search and Genetic Algorithms are adapted to this problem, used and compared. The skills optimized in this work were those responsible for getting the robot to stand after a fall, side-walking and rotating around a fixed point. They are optimized by each algorithm and compared to the original, unoptimized, skills as well as to those of other teams participating in the RoboCup simulated 3D league, where applicable. The achieved results are good, providing skills that performed, according to their specific performance measure such as time elapsed, speed and precision, much better than the original skills, additionally comparing favorably with other teams in the RoboCup 3D simulated league.