Where's WALL-E? A Comparison of the Extended Kalman Filter and Hybrid Inference for Pose Estimation in MAVs

Abstract

Pose estimation is a core competence for cyber-physical systems and is all the more important where there is any element of autonomy. In the context of Micro Air Vehicles (MAVs) this task is more challenging due to weight and cost restrictions. These restrictions dictate that MAVs usually have noisy sensors and limited computational capacity. There are many different approaches to solving this problem but the standard approach is to use the Kalman Filter (KF) [1], or it's nonlinear variant the Extended Kalman Filter (EKF) [2], in order to fuse sensor data and provide optimal estimates of state. While the KF is an optimal estimator of linear systems given some assumptions, most systems are non-linear so the EKF is used. In either case the estimates rely on assumptions that may not always hold. This allows room for improvement. This paper implements the newly proposed technique of Hybrid Inference (HI) [3] on a model of an MAV simulated in Gazebo [4] and explores its performance as compared to the EKF which is used as the standard. HI is a framework for combining graphical models, like the KF, with inverse models which are learned with a Recurrent Message Passing Neural Network (MPNN) [5] [6]. This paper evaluates the technique in a more challenging domain than has previously been implemented. It explores the challenges of implementing the technique, analyses its computational performance and discusses its suitability for use at this time with a strong practical focus. The main findings are that it is too challenging to implement correctly to take full advantage of its proposed benefits. And that it is too computationally inefficient in its current form for it to be suitable for use in real time systems with current technology.