With the acceleration of global aging and the rising incidence of stroke, the demand for lower limb rehabilitation has been steadily increasing. Traditional therapeutic methods can no longer meet the growing needs, which has led to the widespread application of robotic solutions to address labor shortages. In this context, this paper presents a hierarchical and distributed control system based on ROS 2 and Micro-ROS. The distributed architecture decouples functional modules, improving system maintainability and supporting modular upgrades. The control system consists of a three-layer structure, including a high-level controller, Jetson Nano, for gait data processing and advanced command generation; a middle-layer controller, ESP32-S3, for sensor data fusion and inter-layer communication bridging; and a low-level controller, STM32F405, for field-oriented control to drive the motors along a predefined trajectory. Experimental validation in both early and late rehabilitation stages demonstrates the system’s ability to achieve accurate trajectory tracking. In the early rehabilitation stage, the maximum root mean square error of the joint motors is 1.143°; in the later rehabilitation stage, the maximum root mean square error of the joint motors is 1.833°, confirming the robustness of the control system. Additionally, the hierarchical and distributed architecture ensures maintainability and facilitates future upgrades. This paper provides a feasible control scheme for the next generation of lower limb rehabilitation robots.

Introduction: Surface electromyography (sEMG) signals are a key modality for gait phase recognition, as they reflect neuromuscular activation and capture gait dynamics essential for natural control in rehabilitation robotics and human–machine systems. However, reliable phase recognition under complex locomotion conditions remains challenging due to high inter-subject variability, ambiguous phase boundaries, and increased biomechanical diversity across gait types. Objective: To address this issue, this paper proposes a hybrid deep learning framework that leverages the complementary strengths of convolutional neural networks (CNNs) for local time-series feature extraction and bidirectional long short-term memory networks (BiLSTMs) for modeling bidirectional long-range dependencies. Methods: A dataset was collected from 11 healthy subjects performing five representative locomotion scenarios, with gait phases annotated using synchronized motion capture and plantar pressure signals. Results: The proposed method achieved an average classification accuracy of 95.81%, with performance variation across gait types within 4%. The corresponding averaged precision, recall, and F1-score were 94.35%, 95.66%, and 94.95%, respectively, indicating stable and well-balanced recognition performance across multiple locomotion scenarios. Notably, the framework maintained robust performance in biomechanically challenging conditions such as slope ascent and stair ambulation, demonstrating its effectiveness under complex gait dynamics. Conclusion: Overall, these results demonstrate that the proposed framework provides a stable and generalizable solution for sEMG-based gait phase recognition in complex locomotion environments.