Convergent Pose And Twist Estimation For Velocity-denied Mobile Robots Based On A Cascading, Dual-frame, Motion-tracking Estimator

Abstract

Advancements in micro-fabrication and micro-electromechanical systems, increased market demand, and economies of scale have lowered the cost for global navigation satellite system (GNSS) and inertial measurement unit (IMU) sensor systems to levels cost-relevant for average consumers. The combination of GNSS+IMU can provide basic robot localization information, but cannot measure linear velocity, which is essential for autonomous mobile robot operation. Unfortunately, linear velocity sensors like wheeled odometers, air speed, optical flow, or doppler velocity log sensors are situationally applicable and/or cost prohibitive for many robot applications; these robots can be entitled “velocity-denied”. In this work I propose a state estimation algorithm based on a cascading, dual-frame, motion-tracking estimator that is capable of providing accurate pose (positions) and twist (velocities) estimates for velocity-denied robot platforms, by probabilistically estimating the unobserved velocity state based on the time-varying information extracted from the measured position and acceleration states. Because this state estimator is based on a kinematic, motion-tracking state-transition model, it does not require dynamical information about the robot platform or the forces acting on it. I first demonstrate this state estimator algorithm on simulated mobile robot data and then on real data collected from a GNSS+IMU robot sensor system. I show that this state estimation algorithm consistently maintains a dead-reckoning pose accuracy of <1 m of the post-interpolated pose measurements, and provides hidden a linear velocity accuracy of <±1 m/s.