On Behavior Evolution of Legged Robot through Knowledge Array Network

智能复合运动控制（ICMC）是实现智能机器人自主、自适应地完成复杂动作的一种方法。本研究的目的是展示基于ICMC的“基于知识阵列网络的机器人行为进化”的有效性和巨大潜力。以四足机器人TITAN-VIII为例，采用ICMC算法实现了机器人足球射门行为的进化，并验证了该方法的有效性：以关节旋转运动为基本运动，通过重复最优运动组合，逐步学习PTP运动、摆腿运动、跨步运动、靠近运动和踢球运动，得到相应的知识阵列。然后通过知识阵列的层次连接，构建知识阵列网络。最后，从接近和踢球两个方面实现了射门行为。当球的位置已知时，机器人利用知识阵列网络获得的次优控制参数自适应地实现理想的投篮行为，另一方面引入了“复杂度”和“难度”等定量评价机器人行为的指标，以研究有效的行为学习和进化。将其应用于机器人足球行为，验证了指标的充分性，研究了指标值与有效行为学习的关系，阐明了ICMC作为大系统自适应优化方法的特点，这表明了作为一种进化设计方法的巨大潜力，不仅可以应用于机器人的行为学习和进化，而且还可以应用于各种各样的机器人，大型工程系统设计问题。

The Intelligent Composite Motion Control (ICMC) is a method to realize intelligent robots that perform complex actions autonomously and adaptively. The purpose of the research is to show the effectiveness and great potential of "the robot behavior evolution through knowledge array network" based on the ICMC. The Shoot behavior in robot soccer by a quadrupedal robot TITAN-VIII was realized evolutionary based on the ICMC, and the effectiveness of the method was verified.By repeating optimal motion compositions with adopting Joint-Rotation motions as base motions, PTP motion, Swing-Leg motion, Step motion, Approach motion and Kick motion are step-by-step learned and corresponding knowledge arrays are obtained. The knowledge array network was then constructed by hierarchically connecting the knowledge arrays. And finally Shoot behavior was realized from Approach and Kick. If the ball position is given the robot adaptively realizes the desirable shoot behavior by use of the suboptimal control parameters obtained from the knowledge array network.On the other hand several indices for quantitative evaluation of robot behaviors such as "complexity" and "difficulty" are introduced in order to investigate effective behavior learning and evolution. The adequacy of the indices are verified by applying them to the robot soccer behavior, and relation between the index values and effective behavior learning was studied.In addition the characteristics of the ICMC as an adaptive optimization method for large scale systems was clarified, which showed the great potential as an evolutionary design method that can applied not only to robot behavior learning and evolution but also to a wide variety of large-scale engineering system design problems.