毕业生知识库

赋予机器人高效快速的跳跃能力是移动机器人研究中的一个重要课题，部署有跳跃功能的机器人在复杂环境下可以实现越障或避障，能够在更大范围的场景中执行特定任务，因此在灾难救援、野外巡检、室内服务、物品运输、深空探索等方面具有前途广阔的应用价值。但是跳跃运动本身具有高时变性和非线性，同时包含起跳或落地导致的突变特性，合理地为机器人部署跳跃功能是一个难度较高的挑战，具有重要的理论研究价值和实际应用价值。本文面向四足机器人的前向跳跃功能，围绕机器人的仿生结构设计和系统建模、运动轨迹生成和关节轨迹跟踪控制、以及端到端赋能跳跃能力等方面开展研究。本文的主要内容和贡献有以下几个方面：

1. 基于青蛙的运动机制和身体形态设计了一种仿生四足机器人系统的结构框架，并建立了相应的运动学和动力学模型。首先分析了青蛙的跳跃运动机理并构建了包括弹性元件的平面六连杆闭链结构四足机器人模型，利用坐标变换建立机器人各关节和连杆之间的运动学方程。然后对机器人系统各连杆的能量关系进行了分析，求解对应系统哈密顿量的能量函数。接下来运用端口哈密顿方法建立了系统的动力学模型，将机器人关节转轴摩擦和足部与地面静摩擦力作为耗散形式添加到模型中，进而得到了包括系统无源性和耗散性的状态空间表达式。

2. 提出了一种基于非线性优化的机器人关节运动轨迹生成方法。针对基于模型的关节运动轨迹获取问题，首先基于Bernstein方法建立了轨迹的多项式曲线，同时考虑机器人运动时关节的变化特点，构造了Bolza型的跳跃运动目标优化函数，该优化函数包括了机器人从跳跃起始位置到终止位置的运动误差以及广义速度和地面接触力的运行损失。然后运用几何及物理方法设计了相应的约束项避免机器人机身和前后肢之间发生碰撞。使用内点法对设计的优化框架进行优化，得到一组满足各项约束的最优关节运动轨迹，并以实际青蛙跳跃时关节的轨迹变化为例对其合理性进行了类比分析。

3. 提出了一种基于互连与阻尼配置方法(IDA-PBC)的机器人跳跃无源轨迹跟踪控制器。首先分别设计了机器人在起跳阶段、腾空阶段以及落地阶段的运动和稳定判据，以所提出机器人动力学模型为基础，对跳跃过程中不同形式机械能之间的转化关系进行了分析，并据此构建了系统的哈密顿函数。然后通过扩展平衡点集的概念，将IDA-PBC方法扩展为适用于轨迹跟踪控制的形式。建立相应的匹配方程，重新分配了系统的互连矩阵和阻尼矩阵，利用状态反馈方法求解出无源轨迹跟踪控制律，并给出了系统的稳定性分析。最后分别以非线性优化生成的跳跃关节轨迹和实体青蛙跳跃时的关节轨迹为参考轨迹，对机器人跳跃运动时的各关节平滑跟踪效果进行验证，进而证实了所提出的跳跃轨迹跟踪控制器的有效性。

4. 提出了一种基于深度神经网络和强化学习的端到端机器人跳跃技能学习方法。首先将整体的学习框架建立在部分可观测马尔可夫决策过程上，以起到减少传感器的数量和类型，以及降低观测空间维数的作用。强化学习算法使用具有样本高效性的Soft Actor-Critic (SAC)方法进行构建，划分不同跳跃阶段使用不同的模块进行训练。设计了机器人跳跃相关的回报函数，针对不同跳跃阶段的不同特点给出了相应的奖励和惩罚来评判机器人训练中的动作效果。同时引入了好奇心机制和目标导向因子来提高对观测空间和动作空间的探索和利用。训练完成后，本方法能以较高的效率完成满足期望性能指标的四足机器人前向跳跃任务。

Endowing effective and rapid jumping ability to robots is an important topic in mobile robot research. Robots deploying jumping function can realize obstacle sur-mounting or avoidance in complex environments, and execute specific task in the scene with wider range. It also has broad application value in disaster rescue, field inspection, indoor service, goods transportation, deep space exploration, etc. However, jumping locomotion has high time-varying and nonlinear characteristics, and includes sudden changes caused by take-off and landing. It is a difficult challenge for deploying jump-ing function to robots reasonably. Thus, this task has important research meaning and implementation value. This thesis focuses on the basic jumping ability of quadruped robot, and studies bio-inspired structure design, system modeling, motion trajectory generation, joint trajectory tracking control, and end-to-end jumping skill learning. The main works and contributions of this thesis are summarized as follows:

1. Based on the motion mechanism and body morphology, a structure framework of bio-inspired quadruped robot is designed. First, the jumping motion mechanism of frog is analyzed, and a quadruped robot model including elastic components with planar six-linkage and closed-chain structure is constructed. Coordinate transformation is used to build kinematic equations between all the joints and linkages of the robot. Then, energy relationships of different linkages are analyzed, and corresponding energy function with the form of Hamiltonian is solved. System dynamic model is obtained by applying port-Hamiltonian method. Frictions between joint shafts and static friction between ground and the feet of the robot are added into the system model as dissipation.  Thereby, state space expression including the passivity and dissipation of the system is obtained.

2. A joint motion trajectory generation method based on nonlinear optimization is proposed. Aiming at the problem of obtaining model-based joint motion trajectory, firstly, polynomial curves of trajectory are constructed based on Bernstein method. At the same time, considering the property of joint change of the robot during motion, jumping locomotion objective function with Bolza type is designed. Motion errors from the initial position to the final position of the robot, and running cost of generalized velocities and ground reaction forces are both included in objective function. Then, geometric and physical methods are used to design constraints for avoiding collision between the body as well as fore- and back- limbs. At last, interior point method is used to optimize the objective function. A series of optimal joint motion trajectories which satisfy all the constraints are obtained. Comparison analysis of the generated trajectories and jumping trajectories of frog in vivo is given.

3. Based on interconnection and damping assignment, passivity-based control (IDA-PBC), a passive-based jumping trajectory tracking controller of the robot is pro-posed. First, motion and stability criteria of take-off phase, flight phase and landing phase are designed respectively. Conversion relationships of mechanical energy in diff-erent forms is analyzed during the jumping process on the basis of the proposed dyna-mic model, and system Hamiltonian functional is constructed. Then, the concept of equilibrium point set is extended to desired equilibrium trajectory set, and IDA-PBC with trajectory version is obtained. Next, the corresponding matching equation is established, as well as system interconnection and damping matrices are reassigned. The passivity-based trajectory tracking control law is solved by combing state-feedback method, and system stability analysis is given. At last, utilizing the trajectories by optimization generation and by the frog in vivo as references, smooth tracking results of all the joint while jumping are realized. The effectiveness of proposed controller is verified.

4. An end-to-end skill learning method based on deep neural networks and rein-forcement learning for a quadruped robot is presented. First, the overall learning frame-work is established on partially observable Markov decision process. It would play important effects on reducing the number and types of sensors and decreasing the dimension of observable space. Reinforcement learning algorithm is built by soft actor-critic (SAC) with high sample efficiency. Divided jumping phases would be trained by corresponding learning modules. Then, reward functions which are related to robot jumping are designed. According to the different characteristics of different jumping phases, corresponding expectations and penalties are given to evaluate the action effects in training process. At the same time, curiosity mechanism and target-guided factor are introduced to enhance exploration and exploitation of observable space and action space. After training, this method can complete forward jumping task of quadruped robot with desired performance indices effectively and efficiently.