This study introduces a novel non-linear real-time filter based on the concept of Free Energy, applied to estimate the state of mobile robots in dynamic environments. Leveraging Bayesian inference principles, the proposed Free Energy Filter minimizes prediction errors between observed and estimated trajectories, incorporating the Fokker-Planck equation. Simulation results demonstrate effectiveness in improving control accuracy of unicycle robots equipped with the Bessel controller, with comparisons against traditional Kalman filtering highlighting advantages in noisy and uncertain conditions.