AI-based multiple obstacle detection and avoidance for a 6DOF robot manipulator

Abstract

With the rapid advancement of Industry 4.0 in the industrial sector, robotic arms have become crucial in warehouse logistics. Due to the complex and dynamic changing warehouse environments, it is essential for robotic arms to accurately avoid obstacles while working together. Currently, most obstacle avoidance algorithms use pre-trained methods like RRT and A*, but these algorithms consume a lot of computing resources in continuous spaces and when dealing with complex obstacles. In contrast, reinforcement learning algorithms do not require a complete pre-built environmental model and can learn strategies on their own through interaction with the environment. Additionally, the YOLO algorithm is used for fast obstacle detection and, when combined with depth cameras, can accurately determine the specific positions of obstacles. Although LiDAR algorithms offer higher precision, they require more computing resources and are more costly. The YOLO algorithm can balance accuracy and speed in warehouse obstacle avoidance settings. This study aims to improve the autonomous obstacle avoidance capabilities of robotic arms in complex warehouse environments. We combined deep sensors with a retrained YOLO algorithm to accurately detect the positions of multiple obstacles, achieving a high detection accuracy a high confidence level. We used the PPO reinforcement learning algorithm to train the robotic arm for obstacle avoidance, enabling it to effectively navigate around obstacles in complex settings. This research provides an effective solution for autonomous navigation of robotic arms in static complex environments. Future research will further extend to obstacle detection and avoidance in dynamic complex environments.