【作者】 柯文德；

【导师】 崔刚； 洪炳镕；

【作者基本信息】 哈尔滨工业大学 ， 计算机应用技术， 2013， 博士

【摘要】 仿人机器人的机械连杆结构在空间分布上类似于人体躯干与四肢骨骼结构，具备拟人化的双足运动形态与双手处理任务的能力，是人-机器人社会的重要组成部分。仿人机器人运动规划是任务执行与行为决策的基础，其规划方法主要分为基于运动解析方程求解方式与基于人体运动相似性方式。尽管基于运动解析方程的轨迹求解方式具有易于表述与分析的数学特征，并体现出运动轨迹平滑性等优点，但在轨迹自然过渡、复杂动作设计与能量消耗优化上无法与基于人体运动相似性的方法相比。本文在国家863计划重点项目的子项目“竞技与娱乐多机器人系统”课题资助下，研究了基于人体运动相似性的仿人机器人运动规划关键技术，包括：机器人在相似性双足步行中实现平衡与轨迹跟踪控制，在失稳倒地时最小化地面冲击并在倒地后恢复到相似性运动姿势，自主调整固定的人体关节轨迹以适应目标环境改变。主要内容如下：首先，完善了仿人机器人相似性运动系统，从图像捕捉与处理、相似性特征处理、运动约束与优化等方面阐述了相似性运动规划过程，分析了正向运动学与逆向运动学解算以及运动模型简化与重定向设计方法，定义了基于时空控制的相似度，提出了关键姿势判定与提取、运动协调与同步控制以及关键姿势树状结构生成等方法，对运动稳定性判据方法进行了对比，分析了相似性运动中碰撞形式及其原因，阐述了防碰撞约束、落地脚补偿以及关节角度控制等。其次，以7连杆双足步行机器人为例，研究了相似性运动中的平衡与轨迹跟踪控制。分析了双足步行运动学与动力学特性，对相似性运动的关键姿势与基本子相施加运动学、子相衔接与物理条件等约束,建立了基于关节力矩的拉格朗日动力学模型，由机器人雅可比矩阵构建了双足地面接触控制方程，定义仿人机器人动力学模型为连续时间线性时不变系统，引入带观测器的状态反馈控制器以跟踪运动轨迹，从三维倒立摆角度分析步行平衡控制，并由无限时间线性二次型调节器建立运动参数误差最小化目标函数，实现轨迹跟踪控制。再次，研究了仿人机器人相似性失稳倒地时的动作保护与恢复。分析了倒地过程的动力学特性并施加了运动学与物理条件等约束，引入参数化控制方法优化触地点位置，并通过强化技术实现时间变换，将指标泛函求取原时间点的导数转换在新的时间点上求取参数的导数，解决了机器人倒地时触地点位置的寻优问题，使其所受的地面冲击最小、触地位置与倒地稳定性最优；采用可变种群规模的遗传算法对机器人触地后的相似性运动姿势恢复进行了优化。最后，以仿人机器人相似性上台阶运动为例，研究了机器人自主调整固定人体关节轨迹以适应可变目标环境。对机器人相似性上台阶运动过程施加了运动学与物理条件等约束，提出了粒子群分层强化矢量位置择优算法，通过交叉择优操作获得精英粒子，在粒子进化调整策略中选择惯性权重调整策略，并基于嵌入式单目视觉系统采集台阶宽度与高度等参数，用于仿人机器人的膝关节、踝关节轨迹调整上，实现了具有目标环境可变的相似性运动。更多还原

【Abstract】 The spatial mechanical structure of humanoid robot is similar to the skeletalstructure of body and limb of human being, which makes it walk with biped legsand deal with task by hands just like people. For this reason, the humanoid robothas been the important part of human-robot society. The motion planning ofhumanoid robot is the basis of both task executing and behavior deciding. Thereare two planning methods–based on the locomotion analytic equation and basedon the similarity locomotion of human actor. Although there are many advantagesin the method based on locomotion analytic equation such as its easily describ ingand analyzing characteristics of mathematics as well as smooth moving track, thenatural conversion of moving track, complicated motion design and energyconsumption optimization in this method are not as good as that one based on thesimilarity locomotion of human actor.The research is funded by the national863key project "Sports andentertainment multi-robot systems" and it mainly focuses on the key technologiesof motion planning for humanoid robot based on the similarity locomotion ofhuman actor, including keeping balance and following tracks in the biped walkingprocess, minimizing the collision from floor when falling down and recovering tothe similar pose of human actor after it, self adjusting the fixed tracks of humanactor on robot to fit the changeable environment. The main contents are as follow:Firstly, the basic model of similarity locomotion is built up, which isconstructed by three models---image capture and dealing, feature extraction,kinematics constraint and optimization. The forward kinematics and inversekinematics are analyzed as well as the simplfied locomotion model and motionretargeting. The similarity degree based on spatio-temporal control is defined. Themethods of key posture judging and extracting, movement coordinating andsynchronization, and the hierarchy tree structure of key posture are proposed. Thestability judging cristerias are compared. The collisions and causes in similaritylocomotion are analyzied and the corresponding methods such as the collision-freecontraints, landing leg compensation and joint angle control are also described aswell.Secondly, taking the7-link biped walking robot model as an example, thecontrol of keeping balance and following track are studied. The features ofkinametics and dynamics of biped walking are analyzied and the constraints ofkinematics, sub-phase connection and physical condition are set on the key poses and basic phases as well. The Lagrange dynamics equation based on joint torqueare established as well as the biped contacting-land equation through robotJacobian matrix. The dynamics model of humanoid robot is defined as the lineartime invariant system with continous time. The state feedback controller withobserver is used to follow the tracks of human actor. The balance control of bipedwalking is analyzied through the three dimensional inverted pendulum and thetarget function with minimum error of movement parameters is built through thelinear quadratic regulator with infinite time when following tracks.Thirdly, the protection and recovery of pose for humanoid robot are studedwhen the robot falls down. The dynamatics features of falling down for humanoidrobot are analyzed and the contraints of kinematics and physical conditions are setas well. When controlling the contacting position on ground, the parameterizedcontrol method is used to approach the optimal solution and the ehhancingtechnique is use to converse the differential quotient solving with original timepoint through index functional into the parameter acquirement through differentialquotient solving with new time point, which solves the optimizing problem offalling motion for humanoid robot and obtains the minimal floor collision, optimaltouching position and falling stability. After collision with floor, the geneticalgorithm with variable population is used to optimize the pose recovery ofsimilarity locomotion for humanoid robot.Finally, taking the stepping upstairs of humanoid robot as an example, theautonomously fixing tracks obtained from human actor on humanoid robot to fitthe variable target environment is studied. The constraints of kinematics andphysical conditions are set in the process of stepping upstairs for humanoid robot.The particle swarm optimization with hierarchical reinforcement learning isproposed, in which the essence particles are obtained through the crossingselection and the strategies of adjusting inertial weight are selected in the evolvingstrategies of particles. The parameters of stair width and height are obtainedthrough the embedded monocular vision. The tracks of knee and ankle joints onhumanoid robot are adjusted through the above method to fit the changeable targetenvironment.更多还原