Benchmarking and Improving Inverse Kinematics for Endoscopic Inspection Tasks

Abstract

In the context of robotic automation, inspection tasks are becoming increasingly important. In endoscopic inspections, a robot has to enter into a hollow cavity, such as a fuel tank. For motion planning within these constrained environments, inverse kinematics (IK) plays a crucial role. This thesis analyzes the deficiencies of current IK solvers in these environments. For this purpose, a benchmarking software and a benchmarking dataset are created. Based on two state-of-the-art solvers, BioIK and RelaxedIK, several solver improvements with a focus on endoscopic inspections are implemented and evaluated. The results show that directly including collision checking into the solvers is not feasible with the currently integrated collision detection library. The solve rates were significantly increased by implementing cost functions that instead make use of geometric properties of the problem. Additional analyses of the influence of optimization solvers and solver parameters reveal little room for improvement, while adapting the seed state shows greater potential.