Design of Obstacles Avoidance Algorithm for Robotic Arm Using Vision System

Lee, Ching Wern (2025) Design of Obstacles Avoidance Algorithm for Robotic Arm Using Vision System. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text Lee Ching Wern_FULL TEXT.pdf Restricted to Registered users only Download (7MB)

Abstract

Most robotic systems are designed to perform task in a workspace that typically contains many obstacles, such as equipment, humans, and tables. However, these robotic systems are normally not equipped with obstacle avoidance system but solely rely on the emergency stop button or force stop function. An obstacle avoidance system is necessary to prevent the robot from damaging itself or its surrounding objects. To address this, a vision system with depth cameras is applied to gather information in a 3D plane. Two depth cameras was installed at different positions within the workspace to work together, providing the best possible view of the environment. Apart from that, to achieve optimal path planning in both static and dynamic scenarios, the assistance of artificial intelligence is essential. Therefore, a hybrid PF-RL (Hybrid Potential Field and Reinforcement Learning) algorithm is proposed for use in obstacle avoidance path planning. The agent is trained by assigning rewards and penalties based on the distance of the planned path to the goal and the distance between the robotic arm’s end effector and obstacles. To show that hybrid PFRL is an improvement over single potential field method and single reinforcement learning, three algorithms: potential field method, reinforcement learning, and hybrid PF-RL, will be designed and implemented. Performance evaluation was conducted in simulation software like PyBullet, and on real hardware, such as the KUKA KR4 R600 AGILUS. The performance of the algorithm is considered as highly efficient if it achieves the shortest planned path and response time, and requires the adjustment of the smallest end effector angles on the robotic arm while avoiding obstacles. In real world, the integrated obstacles avoidance algorithm in the KUKA robotic arm can be tested by static obstacles like cans, and dynamic obstacles like human hand. The results reveal that PF performs optimally in static environments when the obstacle exists, but it struggles with local minima when the obstacle is far away from the robotic arm, achieving average success rate of 66.67%, compared to 16.67% for RL and 30% for Hybrid PF-RL. Results indicate that the Hybrid PF-RL method significantly outperforms the other two approaches, achieving an average success rate of 96.67% in dynamic environments, compared to 50% for PF and 40% for RL. In dynamic environment, the PF method is computationally efficient but struggles with local minima, while RL handles complex obstacle avoidance but requires extensive training. The Hybrid PF-RL algorithm effectively combines these strengths, resulting in smoother navigation and improved obstacle avoidance in dynamic environments. These findings validate the effectiveness of the proposed approach, making it highly relevant for real-world robotic applications.