Pedro Rafael Duarte

Abstract
The increasing complexity of wireless environments, driven by user mobility and dynamic obstructions, poses significant challenges to maintaining Line-of-Sight (LoS) connectivity. Mobile base stations (gNBs) offer a promising solution by physically relocating to restore or sustain LoS. This paper explores how reinforcement learning (RL) can be applied to gNB mobility control within vision-aided network systems. As part of the CONVERGE project, we present the CONVERGE Chamber Simulator (CC-SIM), a 3D environment for developing, training, and testing gNB mobility control algorithms. CC-SIM models user and obstacle mobility, visual occlusion, and Radio Frequency (RF) propagation while supporting both offline reinforcement learning and real-time validation through integration with OpenAirInterface (OAI). Leveraging CC-SIM, we trained a Deep Q-Network (DQN) agent that proactively repositions gNBs under dynamic conditions. Across three representative use cases, the agent reduced LoS blockage by up to 42% compared to static deployments, highlighting the potential of RL-driven mobility control and positioning CC-SIM as a practical platform for advancing adaptive, next-generation wireless networks.