Practical Vision-Based Monte Carlo Localization on a Legged Robot

Abstract

Abstract-Mobile robot localization, the ability of a robot to determine its global position and orientation, continues to be a major research focus in robotics. In most past cases, such localization has been studied on wheeled robots with rangefinding sensors such as sonar or lasers. In this paper, we consider the more challenging scenario of a legged robot localizing with a limited field-of-view camera as its primary sensory input. We begin with a baseline implementation adapted from the literature that provides a reasonable level of competence, but that exhibits some weaknesses in real-world tests. We propose a series of practical enhancements designed to improve the robot's sensory and actuator models that enable our robots to achieve a 50% improvement in localization accuracy over the baseline implementation. We go on to demonstrate how the accuracy improvement is even more dramatic when the robot is subjected to large unmodeled movements. These enhancements are each individually straightforward, but together they provide a roadmap for avoiding potential pitfalls when implementing Monte Carlo Localization on vision-based and/or legged robots.

An appearance-based visual compass for mobile robots

“…Of course, these methods are based on completely different principles, relying on the recognition of a number of artificial landmarks that were placed around the field to facilitate localization. However, the resulting performance turns out to be comparable: UT Austin [19] did an extensive study on optimizing the localization on the field and reported final accuracies of comparable magnitude.…”

“…The posterior probability distribution that a compass sensor (given a cylindrical map m) estimates when supplied with a measurement z t is given in (19). This distribution can directly be used as the sensor model:…”

“…Here, we assume basically that the perceived color transition patterns originate from very distant objects, ideally at infinity, such that translational component of the camera can safely be neglegted. The posterior stated in (19) forms the sensor model of the visual compass, by relating the measure feature vectors z t to the world state x t , given a cylindrical map m.…”

“…With the sensor model as constructed in (19), it is possible to estimate the robot's yaw angle ψ by learning the cylindrical mapm only at a single training spot. Under the assumption that the transition frequency patterns are produced by objects at infinite distance, it is clearly not possible to estimate the translational component.…”

Metric Localization with Scale-Invariant Visual Features Using a Single Perspective Camera

“…Systems that apply vision-based MCL are also popular in the RoboCup domain. In this scenario, the robots use environment-specific objects as features [19,22].…”