Learning and Dynamic Control of Robot

生物的真实运动是动态的，很难分析。本研究的目的是让机器人通过学习获得运动。我们制作了几个机器人，它们获得了单杠的巨大摆动动作、单腿跳跃运动、两脚行走运动和四个关节的蛇形运动。巨型摇摆机器人巨型摇摆机器人是一个悬挂在单杠上的双摆。在该机器人中，两个摆之间的关节处有一个直流电机，每个关节上都有角度传感器。我们试图通过改变直流电机的力矩来控制巨型摆动。采用基于简单规则库的控制方法，成功地实现了连续巨幅摆动动作。跳跃机器人跳跃机器人有一条来回摆动的腿。摆动滞后的动力由直流电机提供。我们尝试了通过改变直流电机的力矩来控制跳跃运动的方法，类似于1的方法，我们成功地执行了顺序跳跃运动。两足机器人两足步行是一种典型的人类运动。制作了一台具有四个伺服电机的两足机器人，并尝试用神经振荡器网络控制其行走运动。结果，我们成功地完成了稳定的步行动作。四关节机器人下一步，我们制作了四关节机器人，每个关节有四个伺服电机。我们试图让机器人通过学习，利用神经振荡器网络的控制系统来实现高效的蛇行运动。通过学习，我们成功地实现了高效的蛇形运动。在未来的工作中，我们正在考虑让跳跃机器人通过学习来实现跳跃运动。此外，随着这些研究的发展，我们正在考虑用神经网络来控制跳跃运动和双脚行走运动。

The real motion of creature is dynamic and hard to analyze. The purpose of this research is to make a robot acquire the motion by learning.We made several robots, and they acquire giant swing motion of horizontal bar, hopping motion by one leg, two-foot walking motion, and serpentine motion with four joints.(1). Giant Swing RobotGiant swing robot is a double-pendulum hanged on a horizontal bar. In this robot, there is a DC-motor at a joint between the pendulums, and are angle sensors at every joint. We tried to find the way to control the giant swing motion by changing the torque of DC-motor. We succeeded to execute the sequential giant swing motion, by using the control based on simple rule base.(2). Hopping RobotHopping robot has a leg that swings back and forth. The power, which swings the lag, is given by DC-motor. We tried to find the way to control the hopping motion by changing the torque of DC-motor, in the similar way of 1. We succeeded to execute the sequential hopping motion.(3). Two-foot RobotTwo-foot walking is a typical motion of human beings. We made a two-foot robot with four servomotors and tried to control the walking motion with neural oscillator networks. As a result, we succeeded to execute stabilized walking motion.(4). Four-joint RobotTo the next step, we made four-joint robot with four servomotors at every joint. We tried to let the robot get efficient serpentine motion by learning, using control system with neural oscillator networks. We succeeded to execute efficient serpentine motion by learning.For future works, we are thinking to make the hopping robot to achieve the hopping motion by learning.Moreover, as development of those researches, we are thinking to control hopping motion and two-feet walking motion by using neural network.