In this study, we attempt to realize an a utonomous motion planning mechanism of a robot manipulator. Given initial state of a robot manipulator, the appropriate sequence of robot motions to attain the objective state is planned as the solution of the problem. It is difficult to induce a solution of the problem when a complex problem domain is given. We treat this problem under uncertain constraints where unknown obstacles exist in a problem domain. To realize an autonomous reactive planning mechanism, a reinforcement learning algorithm that utilizes trial-and-error processes is applied to this problem. As a reinforcement learning system, we apply genetics-based learning classifier systems. Some numerical experiments are carried. Firstly, to examine the learning performance of the motion planner based on learning classifier systems, simple experiments are accomplished. Next, we give the different task to examine the planning performance of already learned classifier systems for similar tasks. We also carry out another experiment to verify the learning performance of classifier system for more complex problems.
展开▼
