Study on Motor Skill and Autonomous Intelligence in Human and Robot

我们在人类和机器人的运动技能和自主智能方面获得了非常有趣的结果，这将为未来开发家用或宠物机器人做出重大贡献。总结如下。（详情见最终报告）。通过对上肢伸展运动中的运动学习进行实验分析，认为外部环境（黏性力场）可能在大脑中表现为半局部模型。运动学习：运动技能是通过与环境的互动而获得的。我们提出了一种新的程序，通过重复运动来产生适合任务的运动模式。它成功地应用于伸手和击球动作。动态联想记忆模型：提出的联想记忆模型由结构不稳定的振子和化学浓度等共同场组成。结果表明，当给定已知模式时，网络的内部状态为振荡状态；当给定未知模式时，网络的内部状态为混沌振荡状态。在此基础上，提出了一种利用这些动态特性的自主学习算法。计算机模拟表明，额外的学习并没有破坏先前记忆的模式。自主移动机器人的动作学习：将强化学习与递归神经网络相结合，提出了一种新的动作学习体系结构。结果表明，根据任务目标可以获得传感器空间到电机转矩的映射，并在神经网络中间层以抽象表示的形式获得实现目标所需的有用信息。

We obtained very interesting results on motor skill and autonomous intelligence for human and robot, which will give a significant contribution toward developing home or pet robots in the future. The summary is as follows. (see the final report fro detail).1. Motor : model The experimental analysis was performed in terms of the motor learning in the reaching motion of upper limb, and then it was suggested that the external environments (force field of viscosity) might be represented as a semi-local model in the brain.2. Motor learning : Motor skill is acquired through the interaction with his environment. We proposed a new procedure to generate the motor pattern suitable for the task through the repetition of motion. It was successfully applied to the reaching and batting motions.3. Dynamical associative memory model : The proposed associative memory model consists of structurally unstable oscillators and a common field such as chemical concentration. It is demonstrated that when a known pattern is given to the network, the internal state of the network results in the oscillatory one, and when an unknown pattern is given to the network, the internal state oscillates chaotically. Further, a new autonomous learning algorithm utilizing these dynamical characteristics was proposed. Computer simulation demonstrated that the additional learning did not destroy the previously memorized patterns.4. Action learning for autonomous mobile robot : We proposed a new architecture for action learning by combining reinforcement learning with recurrent neural network. It was shown that the mapping from the sensor space to the motor torque could be acquired according to the task goal and the useful information to reach the goal could be obtained as an abstract representation in the middle layer of the neural network.