RobustStateNet: Robust ego vehicle state estimation for Autonomous Driving

Abstract

Control of an ego vehicle for Autonomous Driving (AD) requires an accurate definition of its state. Implementation of various model-based Kalman Filtering (KF) techniques for state estimation is prevalent in the literature. These algorithms use measurements from IMU and input signals from steering and wheel encoders for motion prediction with physics-based models, and a Global Navigation Satellite System(GNSS) for global localization. Such methods are widely investigated and majorly focus on increasing the accuracy of the estimation. Ego motion prediction in these approaches does not model the sensor failure modes and assumes completely known dynamics with motion and measurement model noises. In this work, we propose a novel Recurrent Neural Network (RNN) based motion predictor that parallelly models the sensor measurement dynamics and selectively fuses the features to increase the robustness of prediction, in particular in scenarios where we witness sensor failures. This motion predictor is integrated into a KF-like framework, RobustStateNet that takes a global position from the GNSS sensor and updates the predicted state. We demonstrate that the proposed state estimation routine outperforms the Model-Based KF and KalmanNet architecture in terms of estimation accuracy and robustness. The proposed algorithms are validated in the modified NuScenes CAN bus dataset, designed to simulate various types of sensor failures.