In this paper, we present a system that integrates computer vision with machine learning to enable a humanoid robot to accurately solve any 3 dimensional maze that has not been previously given to it. The robot can construct the optimum path policy based on previous iterations and does not require any specialized programming. The experimental setup includes a constructed 3D maze with a start and end point. The robot solves the maze using a red-colored ball. The robot can physically tilt the base of the maze with its hand so that the ball can roll into the desired region. The robot would begin tilting the maze only if a path exists between the start and the end point. If none exists, the robot would remain idle. This work is important and novel for a couple of reasons. The first is to determine if constant repetition of a task leads to gradually increasing performance and eventual mastery of a skill. If yes, can that skill be adapted to a generic ability ( Fleishman, 1972)? Also, can a robot's performance match or exceed that of an average human in the acquired ability?