DEVELOPMENT OF AN OPTIMAL PREDICTIVE COLLISION AVOIDANCE SYSTEM FOR HUMAN-ROBOT COLLABORATION

Abstract

This thesis presents the development of a collision avoidance system for human robot collaboration applications where a human shares its workspace with a robotic arm. A vision system is used to capture colour point clouds of the shared workspace in real-time. An easy-to-use calibration method is then developed to transform the point clouds to the robot’s coordinate frame. Algorithms for processing the transformed points to obtain models of the human and static obstacle(s), and a method used to model the robot, were developed next. The human is modelled by a plane in front of their torso, and any body parts or objects in the front of this plane are modelled by spheres. This model is more computationally efficient compared to the spheres-only modelling method. Methods for predicting the motions of the human’s arms, head and torso were also developed. Next, two predictive collision avoidance algorithms that account for the system dead time were developed. Five metrics were proposed for evaluating the collision avoidance and trajectory tracking performances of the algorithms. The 1st collision avoidance algorithm is an optimal algorithm, and the 2nd collision avoidance is a non-optimal algorithm. The robotic arm used in this research is an Elfin 5 industrial robot. Simulations of the robot avoiding a static obstacle using the 2nd algorithm demonstrated the necessity of using prediction whenever system dead time is present. Both algorithms were then simulated with a human model moving with a constant speed blocking the robot’s path. Based on the superior performance of the 1st algorithm in collision avoidance and trajectory tracking, it was selected for the rest of the collision avoidance experiments. Three real-time human robot collaboration scenarios requiring collision avoidance were studied experimentally. The experimental results demonstrate the consistent excellent collision avoidance and trajectory tracking performance of the 1st algorithm in all scenarios. The algorithm accelerates the robot to compensate for the time spent performing collision avoidance so that no productivity is lost, unlike many prior approaches. Its computational speed is also faster than other optimal collision avoidance algorithms for human-robot collaboration.