強化学習を用いた分散学習エージェントによるマニピュレータのリアクティブプランニング

概要

論文の詳細を見る
In engineering applications such as manipulator motion planning in an unknown space, we have to achieve the following requirements : planning in dynamically changing environments including mobile obstacles and even breakdown of some parts of the manipulator. Reactive planning is one of the methods that can overcome the problems, because it decides an action for each state reactively. For giving more flexibility to the planning mechanism, a distributed implementation will be effective. In this paper, we attempt to construct a reactive planning mechanism composed of distributed learning agents. The learning agent is assigned to each joint of a manipulator. It observes a condition around the joint and decides an output torque of the joint probabilistically. Then, it updates the action selection probability according to an evaluation using a reinforcement learning scheme. Finally, some computational simulations verify the behavior of the proposed mechanism.
一般社団法人日本機械学会の論文
1995-01-25