deep learning supported obstacle avoidance for universal ur10 collaborative robotic arm

Abstract

This Thesis presents a software framework that utilizes copper foil electrodes and an FDC2114 LC-resonant capacitance-to-digital converter to mathematically model the detection of conductive objects within a 30 cm radius of the cobot. The Soft Actor-Critic (SAC) DRL algorithm was trained through a four-stage curriculum within the CoppeliaSim simulation software, using a 61-dimensional state space that includes different measurements of the cobot’s sensors and its joint angles. A Control Barrier Function (CBF) prevents the cobot from entering a state that does not satisfy the safety constraint of remaining at a minimum distance to the detected obstacles, and a dual-loop Lagrangian method ensures the cobot reaches its goal while simultaneously satisfying the CBF constraint.