Framework for Filtering Neural Network Predictions for Robust State Estimation

Learning-enabled components (LECs) are pervasive in the field of autonomous systems, especially for perception. Despite how common they are, there is a lack of a fundamental method for managing their uncertainty and estimation error. In this work, we advocate for an approach for training LECs for perception that admits a principled model of perception error. Further, we propose how tools from estimation theory and filtering may be adapted to compensate for these errors for reliable operation of closed-loop control with perception LECs.