Gaussian process-enhanced control barrier functions for safe obstacle avoidance on a UR5 manipulator

This thesis develops a learning-based Control Barrier Function (GP-CBF) safety filter for a 6-DOF UR5 manipulator, operating in ROS 2 and Gazebo, that protects the full robot. The goal is to execute nominal tasks while avoiding collisions with static and moving obstacles, even when obstacle geometry is not known a priori. We first implement a CBF safety filter for the end-effector of a 6-DOF UR5 manipulator and then we develop: (i) an end effector GP-CBF that learns online a safety function from obstacle surface samples and enforces it through a single linear CBF constraint, and (ii) a model-based CBF-QP for whole-body safety with known obstacles that protects multiple points distributed on the links (with an optional ground-plane constraint). Merging these two implementations, the main contribution of this thesis is a GP-CBF safety filter for the full manipulator : the GP posterior mean at protected points defines the barrier values, its spatial gradient defines the sensitivities, and a quadratic program in joint velocities minimally deviates from the nominal command whenever the learned CBF constraints activate. This merges the end-effector GP learning with the multi-constraint wholebody CBF, so that link and ground clearances are enforced even with unknown obstacles. The complete pipeline includes nominal Cartesian control mapped to joints via damped least squares, online GP inference with Squared Exponential and Reciprocal Distance kernels, and a CBF-QP that guarantees nonnegative barrier certificates when applied. The simulations performed show that the unified GP-CBF filter preserves task performance while maintaining safety margins for the end effector and all protected link points in scenarios with single or multiple obstacles, including moving ones.