Sample-efficient learning of soft priorities for safe control with constrained Bayesian optimization

A complex motion can be achieved by executing multiple tasks simultaneously, where the key is tuning the task priorities. Generally, task priorities are predefined manually. In order to generate task priorities automatically, different frameworks have been proposed. In this paper, we employed a black-box optimization method, i.e. a variant of constrained Bayesian optimization to learn the soft task priorities, guaranteeing that the robot motion is optimized with high efficiency and no constraints violations occur during the whole learning process.