Vision-Based Obstacle Avoidance Strategy for Motion Control of Robotic Arm

Loke, Guan Yan (2026) Vision-Based Obstacle Avoidance Strategy for Motion Control of Robotic Arm. Final Year Project (Bachelor), Tunku Abdul Rahman University of Management and Technology.

Text LOKE GUAN YAN_Full Text.pdf Restricted to Registered users only Download (32MB)

Abstract

This research presents a vision-based robot arm motion control approach to obstacle avoidance that solves the limitations of traditional pre-programmed systems. Traditional robotic arms are limited by their inability to adapt to changing environments. In dynamic settings, these systems rely on preprogrammed paths, requiring manual intervention from engineers or operators to adjust for new obstacles. This causes inefficient work, increased chances of human error, and potential collisions, raising significant safety concerns. The lack of real-time adaptability limits the versatility of robotic arms, which is crucial for applications in industries that demand continuous operation without frequent reprogramming. This study successfully developed and validated a Neural Network Enhanced-Bidirectional RRT* (NNE-BiRRT*) algorithm that integrates Bidirectional RRT* (BiRRT*) path planning with a neural network for sampling node prediction and an inverse kinematics solver using another separate neural network to optimise efficiency. Three RGB cameras and YOLO11 object detection algorithm provide real-time localisation of obstacles, with a third neural network used to find the 3D location of objects. The validation was tested across 130 individual test scenarios in a PyBullet simulation with a 6 degree-of-freedom (DOF) FANUC M-710iC robot arm for static and dynamic obstacle scenarios. Evaluation criteria are the execution time, computational cost by counting the generated nodes, and the length of path compared to established algorithms like RRT* and BiRRT*. Tested scenarios included 1,2, and 3 obstacles for static case and 2 obstacles for dynamic case. Results demonstrated significant improvements over traditional approaches. The NNE-BiRRT* algorithm achieved up to 94% reduction in computational nodes, decreased path generation times from up to 13.5 seconds to under 1 second, and maintained optimal path quality with shorter path lengths. The inverse kinematics neural network provided consistent computation times of 1.26-2.27 seconds compared to 3.84-13.02 seconds for traditional Jacobian methods, enabling real-time applications. The vision system achieved coordinate prediction accuracy below 1% error, which was tested with different obstacle arrangements. This research contributes significantly to safer human-robot collaboration, enhanced industrial automation efficiency, and advances toward Industry 4.0 smart manufacturing, aligning with Sustainable Development Goal 9 (Industry, Innovation, and Infrastructure) through improved operational efficiency, added safety features, and equipment lifecycle extension