【作者】 王晓林；

【导师】 臧希喆；

【作者基本信息】 哈尔滨工业大学 ， 机械电子工程， 2012， 硕士

【摘要】 被动动力行走是机器人领域一个重要的研究方向，重点是对步行机理、步态特性的分析，同时结合其行走步态高效、稳定、自然、柔顺的优势，在对机器人自身结构及动力学优势充分发掘的基础上，为机器人样机的建造和推进实用化提供新的思路。本文将建立双足有膝关节机器人动力学模型，并围绕轨道稳定性和全局稳定性分析被动机器人的步态特性，进一步了解被动行走本质。同时为提高机器人适应能力，提出基于强化学习的被动行走自适应控制策略，并通过仿真和样机实验验证控制策略的有效性。首先，在对被动机器人模型演化过程和直腿点足模型运动特点分析的基础上，建立一种带膝关节点足简化模型。结合人类行走特点对模型进行步态划分，通过理想化假设，采用Lagrange方程和角动量守恒推导摆动阶段和碰撞阶段的动力学方程，并给出数值仿真流程。在此基础上，建立混合系统数学模型，基于庞加莱映射原理，采用牛顿-拉普森迭代法求解不动点，分析步态误差同Floquet乘数关系；以求得不动点为初值，深入分析稳定被动动力行走步态特性。利用胞映射法，求解膝关节模型行走步态收敛域范围，并改变胞元密度以观察步态收敛域变化。针对纯被动模型步态收敛域范围小、稳定性差的缺陷，在保证纯被动行走优势的前提下，在髋关节添加PD控制使之成为准被动形式，并分析控制器参数变化对行走步态产生的影响。为模拟控制器效果，在Adams环境下建立模型进行动力学仿真，并对仿真结果进行分析。然后，在以PD控制作为底层控制的基础上，建立基于CMAC网络的Sarsa(λ)增强学习控制策略，对增益参数进行优化，提高机器人行走自适应性。应用该控制策略，在不同倾角的斜坡平面上进行仿真，对控制效果进行分析。最后，在实验室已有机器人样机的基础上进行结构和硬件改进，移植控制算法到实验样机中，并进行了不同坡度下的行走样机实验。通过对实验结果的对比分析，验证控制策略的有效性。更多还原

【Abstract】 Passive dynamic walking is an important research area in the field of roboticsfocusing on the analysis of the walking mechanism and gait characteristics. It canprovide new ideas to construct the prototype of robot and promote the practical usebecause of its efficient, stable and natural gait, as well as the advantages of itsmechanical structural. Here we establish the model of robot with knees and study itsnature of passive dynamic walking by analyzing orbit stability and global stability.In order to enhance the adaptability of the robot, reinforcement learning controlstrategy is also proposed and verified by numerical simulation and prototypeexperiment.Firstly, by analyzing the evolution and characteristics of the existed models, webuilt a new model with point feet and straight legs. In order to deduce the swingphase and collision phase equation, Lagrange method and angular momentumconservation principle was used. We divide walking phase by referring to thecharacteristic of human gait, then numerical simulation process is given.Secondly, the mathematical hybrid system of the model is built. The Newton-Raphson iterative method for solving fixed point is given based on the Poincarémapping principle to analyze the relationship between Floquet multiplier and gaitdeviation. The fixed point is set as initial value to analyze the characteristic forpassive dynamic walking. We use cell mapping method to get the range of domain ofattraction, and change the cell density to observe the variation.For the domain of attraction is small and gait is unstable, PD controller is addedat hip under the premise of guaranteeing its advantage. In order to verify the controleffect, model is built in Adams and simulation result is analyzed.Then, PD controller is set as the underlying control, and we build the Sarsa(λ)reinforcement learning control strategy based on CMAC network to enhance therobot adaptability, simulation results are analyzed in different inclination of gravityfield.Finally, rationalize the structure and hardware of the prototype, and thereinforcement learning control strategy has been transplanted to the robot, so the effectiveness of the algorithm can be verified by the experiment.更多还原