Non-contact control technologies hold immense promise in industries, yet achieving agile and efficient non-contact manipulation in high-dimensional spaces remains a formidable challenge. This study introduces a novel deep reinforcement learning (DRL)-driven magnetically levitated Delta-robot system, termed the Maglev-Delta robot. Theoretically, we delineate the fundamental prerequisites for magnetic levitation control and propose an optimized magnet array configuration to maximize the controllable domain, thereby enabling the design of a high-dimensional levitation control execution module. A method for nonlinear attenuation of magnetic field strength is introduced to address the issue of actuator entrapment near the magnets, which leads to a scarcity of training samples for the DRL controller. Additionally, we construct a novel reward function model balancing movement speed and levitation precision to enhance the levitation control performance of the DRL controller. Experimental results demonstrate that the developed Maglev-Delta robot can achieve high-speed and high-precision two-dimensional and three-dimensional levitation control tasks, showcasing exceptional flexibility. Notably, in simulated handling tasks, the robot was able to stably complete load handling tasks. Based on the experimental results, we analyze that the scaled-up Maglev-Delta robot can execute levitation maneuvers within a substantial 27×27×27 m3 volume, capable of manipulating masses up to 3.8×10? kg, thereby underscoring its vast potential for practical applications.