Adaptive Sensor Fusion Using Machine Learning for Robust Localization in Low-Cost Autonomous Robots

Robot sensor data is prone to becoming unreliable due to factors like wheel slippage, inertial errors, reflective surfaces, insufficient lighting, and occlusions. A typical EKF uses constant measurement covariances; therefore, the filter may continue relying on a degraded sensor even when its uncertainty varies. The paper introduces the EKF coupled with the light neural network, which predicts the level of confidence in encoder, inertial, ultrasonic, and visual sensor measurements and recalibrates measurement covariance prior to the EKF update step. A Python simulator produced 100,000 training samples to test dead reckoning, a conventional EKF, an innovation-based adaptive EKF, and the proposed approach in 500 Monte Carlo experiments under normal, mild, and severe sensor degradation. In severe sensor degradation conditions, the conventional EKF reached 743.6 cm RMSE, while the proposed solution constrained RMSE to 251.7 cm, reducing it by 66.2%. The performance of the proposed solution was statistically indistinguishable from that of adaptive EKF in a ±10.0 RMSE range.