Magnetically actuated capsule endoscope enables wireless navigation for minimally invasive diagnosis and monitoring of gastrointestinal (GI) diseases. However, current studies lack reliable solutions for maintaining stable magnetic levitation at arbitrary positions within the stomach for imaging and observation during clinical procedures. This work addresses this challenge by leveraging a reinforcement learning (RL) policy to achieve magnetic levitation of capsule endoscope for path tracking under a single permanent magnet. To minimize the sim-to-real gap and enable effective policy training, we develop a simulation environment based on NVIDIA Isaac Sim. In this setup, a permanent magnet is mounted on the end-effector of a robotic arm, enabling position control of the capsule robot using magnetic gradient force. Proximal Policy Optimization (PPO) is selected as the RL algorithm due to its balance of training stability and sample efficiency. The proposed framework is implemented on a robotic platform to achieve path tracking control of a capsule endoscope in a magnetic levitation manner, demonstrating an average positional error of 0.62 mm.