Skip to Main Content
Sensor errors can adversely affect the behavior of a control system. When multiple sensors are used, a broken sensor can have its effects minimized by artificially inflating its error covariance. In this paper, a different approach to compensating for sensor errors in a multiple-sensor control system is introduced. The technique, referred to as a neural extended Kalman filter (NEKF), is developed for closed-loop control systems. The NEKF learns on-line from the same residual information used in the state estimator. The improvement in the sensor report is made by the neural network being added to the measurement model. In this work, the NEKF is applied to vehicle trajectory control problem with a position sensor and a velocity sensor.