No abstract

Abstract.We propose an integrated technique of genetic programming (GP) and reinforcement learning (RL) that allows a real robot to execute real-time learning. Our technique does not need a precise simulator because learning is done with a real robot. Moreover, our technique makes it possible to learn optimal actions in real robots. We show the result of an experiment with a real robot AIBO and represents the result which proves proposed technique performs better than traditional Q-learning method.

The cooperation of several robots are needed for complex tasks. The cooperation methods for multiple robots generally require exact goal or sub-goal positions. However, it is difficult to direct the goal or sub-goal positions to multiple robots for the sake of cooperation with each other.Planning algorithms will reduce the burden for this purpose. In this paper, we propose a multi-agent planning algorithm based on a random sampling method. This method doesn't require the exact sub-goal positions nor the times at which cooperation occurs. The effectiveness of this approach is empirically shown by simulation results.

The multi-agent cooperation has been proved useful in executing many complex tasks. In our previous paper, we proposed a path planning algorithm based on a random sampling for the sake of the multi-agent cooperation. However, the action path of the robots is liable to be deviated by the noise in the real world. Thus, some correction mechanism is required to reduce the differences between the planned path and the real one.In this paper, we propose a re-planning method to solve the above-mentioned difficulty. The applicability of this method is confirmed with the experiment using two humanoid robots, in which they have re-generated path plans according to their locations detected by their own cameras.