Robots with Brain-Like Information Processing

本研究的主要成果可概括为以下四部分：(1)多传感器躯干机器人的研制研制了双7自由度三指手躯干机器人，并装配了多传感器装置。即六个三轴指尖力传感器，一个六轴腕力/扭矩传感器，关节角度传感器和双CCD相机，其图像发送到两个不同的图像处理系统。(2)根据反应性行为网络的内部参数，开发了一种学习算法。利用躯干机器人和多感官系统对原始反应行为进行设计和编程。然后将该算法安装到躯干机器人中，并应用于反应抓取。(3)具有双存储记忆模型的类脑信息处理系统。采用神经网络技术设计了具有长期和短期存储的联想记忆模型，并对躯干模型进行了实现。在实验中，每三秒钟向机器人展示并教授四个彩色方块的十个顺序放置，同时重复内部排练以加强记忆。基于联想记忆，机器人用手臂和手指对序列进行回放，体现了人类记忆的特点。(4)用自我运动原素理解他人脑科学认为，模仿是人类和大型灵长类动物符号操作、语言和逻辑思维的基础。我们研究了仿人机器人运动原语的实现。采用概率神经网络和隐马尔可夫模型进行集成。并对其全身运动进行了计算机仿真，验证了该方法的有效性。

The main results of this research are summarized as in the following four paragraphs :(1) Development of a Multi-Sensored Torso RobotA torso robot with dual 7 DOF manipulators with a three-fingered hand each was fabricated and equipped with multi-sensory devices. Namely, six of three-axes finger-tip force sensors, a six-axes wrist force/torque Sensor, joint angle Sensors, and dual CCD cameras whose images are sent to two different image processing systems.(2) Grasp With Reactive Behaviors' NetworkA learning algorithm was developed hr the internal parameters of the reactive behavior's network. The primitive reactive behaviors were designed and programmed using the torso robot and the multi-Sensory system. The algorithm was then installed into the torso robot and applied to the reactive grasp.(3) Brain-Like Information Processing System with the Dual-Storage Memory ModelThe associative memory model with the long-term and short-term storage was designed using neural networks and implemented for the torso model. In the experiments, ten sequential placements of four colored blocks were shown and taught to the robot every three seconds while the internal rehearsal was repeated to strengthen the memory. The robot playbacked the sequence with the arm and fingers based on the associative memory, which indicated the features of human memory.(4) Understanding the Others with the Self Motion PrimitivesIt is believed ill the brain science that the mimesis is the basement of symbol manipulation, language, and logical thinking of the human and the large primates. We investigated the implementation of mimesis using the motion primitives of humanoid robots. The probabilistic neural networks and the hidden Markov model were devised for the integration. The computer simulation of the whole body motion were executed and showed the fundamental effectiveness.