Autonomous Obstacle Avoidance System for Spider parallel robot Based on Deep learning visual Perception

Dynamic obstacle avoidance constitutes a critical component in enabling autonomous mobility and safe operation of robots. In complex and dynamically changing operational environments, robots must possess real-time obstacle detection and dynamic path planning capabilities. This study employs a spider parallel robot as the platform, integrating an RGB-D depth camera and an inertial measurement unit to construct an environmental perception system, thereby providing the robot with multimodal information such as three-dimensional visual data and attitude angles. Firstly, an enhanced object detection model, YOLOv4-M, is proposed based on YOLOv4. By incorporating an additional detection layer, introducing a spatial pyramid pooling structure, and optimizing anchor boxes using K-means clustering, the model improves the accuracy and robustness of multi-scale obstacle recognition. The model identifies and localizes obstacles within color images and employs depth map alignment techniques to determine obstacle dimensions and their spatial distances relative to the robot. Based on real-time robot attitude data and obstacle identification results, an improved artificial potential field-based obstacle avoidance decision-making model is developed. This model integrates a dead-zone distance mechanism and a dynamic steering strategy to effectively address local minima issues inherent in traditional potential field methods, thereby enabling dynamic walking path planning. The generated decision commands are transmitted to the robot's motion control unit, ultimately achieving autonomous obstacle avoidance in unknown environments. Both simulation and physical experiments validate that the proposed system effectively accomplishes autonomous obstacle avoidance tasks for the spider parallel robot. This research provides a technical foundation and reference for obstacle recognition and autonomous navigation in robots relying solely on visual and inertial sensors.