Robotic data gathering (RDG) is an emerging paradigm that navigates a robot to harvest data from remote sensors. However, motion planning in this paradigm needs to maximize the RDG efficiency instead of the navigation efficiency, for which the existing motion planning methods become inefficient, as they plan robot trajectories merely according to motion factors. This paper proposes radio map guided model predictive communication (MPCOM), which navigates the robot with both grid and radio maps for shape-aware collision avoidance and communication-aware trajectory generation in a dynamic environment. The proposed MPCOM is able to trade off the time spent on reaching goal, avoiding collision, and improving communication. MPCOM captures high-order signal propagation characteristics using radio maps and incorporates the map-guided communication regularizer to the motion planning block. Experiments in IRSIM and CARLA simulators show that the proposed MPCOM outperforms other benchmarks in both LOS and NLOS cases. Real-world testing based on car-like robots is also provided to demonstrate the effectiveness of MPCOM in indoor environments.