随着我国人口老龄化加剧和人们对机器智能化的不断追求，可以预见，未来机器人将在各类服务场景中部分取代传统的人力劳动。其中，人形机器人外形与人类高度相似，因此在融入复杂人居环境、实现自然流畅的人机交互等问题上被寄予了厚望。面对多样、复杂的任务场景，通用的人形机器人控制框架和可泛化的控制策略是可行的解决方案。本论文从人形机器人多步态行走控制和跳跃运动控制两方面着手，借鉴人脑运动控制机制，提出了一套通用的脑启发式人形机器人腿足运动控制策略。本文的主要研究内容如下：首先，基于Unity仿真环境搭建了可用于算法设计、策略训练和可靠性验证的人形机器人仿真平台。包括了机器人数学建模、运动学分析和多样化训练场景、验证场景搭建等基本组成部分，为算法研究和实验验证提供基础。其次，借鉴大小脑协同控制模式，设计一套通用的人形机器人控制框架。该框架采用模型、数据双驱动的脑启发式控制思路，结合基于运动学模型的前馈控制信号和基于强化学习的反馈控制信号，使得机器人具备一种与具体步态解耦的泛化具身平衡控制能力。随后，论文在该控制框架基础上对人形机器人多步态行走控制问题和跳跃运动控制问题展开研究。针对两种运动的特性，采用课程学习的思想，进行对应的分阶段训练设计。最终经实验验证，多步态行走控制策略实现了多种步态下对身体躯干和左、右脚踝的位姿轨迹跟踪，满足人居环境中机器人位置、姿势变换和避障轨迹规划等多种需求。基于虚拟模型的跳跃运动控制策略利用弹簧负载倒立摆虚拟模型解决了欠驱动条件下双足机器人跳跃运动控制前馈参考轨迹生成问题，实现了不同距离及高度的跳跃运动控制，增强了机器人在不连续地形表面的移动能力。综上所述，本文围绕人形机器人运动控制问题，以人脑控制机制为启发，设计了一种双驱动的通用控制框架，通过运动学分析、强化学习、课程学习等方法，实现了人形机器人的多步态行走运动控制和跳跃运动控制，并在仿真环境中验证了策略的有效性。
As the aging population of our country intensifies and people continue to pursue machine intelligence, it is foreseeable that in the future, robots will partially replace traditional human labor in various service scenarios. Among them, humanoid robots, due to their highly similar appearance to humans, are expected to play a significant role in integrating into complex human environments and achieving natural and smooth human-robot interactions. Facing diverse and complex task scenarios, a general humanoid robot control framework and generalized control strategies are feasible solutions. This thesis focuses on the control of humanoid robot multi-gait walking and jumping movements, drawing inspiration from human brain control mechanisms, and proposes a set of general brain-inspired humanoid robot leg-foot motion control strategies. The main research contents of this thesis are as follows:Firstly, a simulation platform for humanoid robots was built based on the Unity simulation environment, which can be used for algorithm design, strategy training, and reliability verification. It includes basic components such as robot mathematical modeling, kinematic analysis, and the construction of diverse training and validation scenarios, providing a foundation for algorithm research and experimental verification.Secondly, inspired by the cerebellum-cerebral cortex cooperative control mode, a general humanoid robot control framework is designed. This framework adopts a brain-inspired control approach driven by both models and data, combining feedforward control signals based on kinematic models and feedback control signals based on reinforcement learning, enabling the robot to possess a generalized whole-body balance control capability decoupled from specific gaits.Subsequently, the thesis conducts research on multi-gait walking control and jumping motion control of humanoid robots based on this control framework. According to the characteristics of the two movements, a curriculum learning approach is adopted for corresponding staged training design. Finally, through experimental verification, the multi-gait walking control strategy achieves trajectory tracking of body trunk and left and right ankle poses under various gaits, meeting various requirements such as robot position and posture transformation, and obstacle avoidance trajectory planning in human living environments. The jumping motion control strategy based on the virtual model utilizes a spring-loaded inverted pendulum virtual model to solve the problem of feedforward reference trajectory generation for bipedal robot jumping motion control under underactuated conditions, achieving jumping motion control at different distances and heights, enhancing the robot‘s mobility on discontinuous terrain surfaces.In summary, this thesis revolves around the motion control problem of humanoid robots, drawing inspiration from human brain control mechanisms, designing a dual-drive general control framework, and realizing the multi-gait walking motion control and jumping motion control of humanoid robots through methods such as kinematic analysis, reinforcement learning, and curriculum learning, and verifying the effectiveness of the strategies in a simulation environment.
