下肢外骨骼机器人可以被用于增强人体的负重能力、行动能力和耐力，在长途跋涉中起到减缓身体疲劳的作用。下肢外骨骼作为一种贴身穿戴的机电设备，在使用中与人体存在大量的交互。物理性人机交互技术(pHRI)旨在结合人与机器人的优势，使机器人与人体达到更好的协同配合效果，目前在包括工业、医疗等的许多领域中，发挥着重要作用。然而，物理性人机交互技术的安全性和效率尚未得到系统解决。本研究首先从硬件和软件两方面解决外骨骼的安全性问题，硬件方面，采用具有缓冲吸能作用的柔性驱动器--串联弹性驱动器；软件方面，自适应变阻抗控制在算法层面赋予机器人主动柔顺特性。接着，本文针对柔性驱动下肢外骨骼机器人设计了一种自适应变阻抗控制器，其基于一个可变的期望阻抗模型，控制机器人跟踪并辅助人体的运动。在变阻抗控制器的设计方面，本文从不同的考虑因素出发，制定了基于交互力和基于异常检测的两种变阻抗策略。基于交互力的策略根据所要追踪的角度与机器人当前关节角度之间的差值在线调整阻抗参数，使机器人能够在贴合人体的同时为人体留出部分自由运动的空间，从而实现更好的协作效果，从而提高pHRI的安全性、效率和舒适性。基于异常检测的策略从学习的角度出发，利用所构建的神经网络从一系列本体感知传感器接收多模态的信息，并输出表征物理性冲突的异常分数，并利用该异常分数在线调整控制器的阻抗参数，确保安全高效的人机交互。该基于异常检测的方案允许机器人发掘潜藏在人机交互过程中的语义信息(例如步态相位不匹配，身体失衡，身体疲劳)，从而对人机间的冲突做出适当的反应。最后，本文利用李雅普诺夫方法对闭环系统的稳定性进行了严格的验证，并将算法部署在下肢外骨骼上以测试其有效性，并基于实验结果对所设计控制器的性能进行了评估。

The lower-limb exoskeleton robot can be used to enhance the mobility and endurance of the human body, and contribute to reducing human fatigue during long-distance walking. There is plenty of physical interaction between the lower-limb exoskeleton and the human body during use. Physical Human-Robot Interaction(pHRI) aims to combine the advantages of humans and robots to achieve a better synergistic effect between robots and humans. It is currently playing an important role in many fields including industry and medical care. However, the safety and efficiency of pHRI have not been systematically addressed. This research first solves the safety problem of exoskeleton from two aspects of hardware and software. In terms of hardware, compliant actuators with the advantages of elasticity and energy-absorbing are used; in terms of software, adaptive variable impedance control endows the robot with active compliance characteristics at the algorithm level. Next, this thesis designs an adaptive variable impedance controller for a compliant-driven lower-limb exoskeleton robot, which controls the robot to track and assist the movement of the human body based on a variable desired impedance model. Regarding the design of variable impedance controllers, two variable impedance strategies based on interaction force and anomaly detection are formulated based on different considerations. The strategy based on interaction force adjusts the impedance parameters online according to the difference between the desired angle and the current joint angle of the robot, so that the robot can fit the human body while leaving some free movement space for the human body, so as to achieve better pHRI, thereby increasing the safety, efficiency, and comfort of pHRIs. From the perspective of learning, the strategy based on anomaly detection uses the constructed neural network to receive multi-modal information from a series of proprioception sensors, outputs anomaly scores representing physical conflicts, and uses the anomaly scores to adjust the controller‘s impedance parameters to ensure safe and efficient human-computer interaction. This anomaly detection-based scheme allows the robot to explore the semantic information (e.g., gait phase mismatch, body imbalance, body fatigue) hidden in the human-robot interaction, so as to respond appropriately to human-robot conflicts. Finally, the stability of the closed-loop system is rigorously verified using the Lyapunov method, and the algorithm is deployed on a lower-limb exoskeleton to validate its effectiveness, and the performance of the designed controller is evaluated based on the experimental results.