Statistical Study of the Performance of Recursive Bayesian Filters with Abnormal Observations from Range Sensors

Abstract

Range sensors are currently present in countless applications related to perception of the environment. In mobile robots, these devices constitute a key part of the sensory apparatus and enable essential operations, that are often addressed by applying methods grounded on probabilistic frameworks such as Bayesian filters. Unfortunately, modern mobile robots have to navigate within challenging environments from the perspective of their sensory devices, getting abnormal observations (e.g., biased, missing, etc.) that may compromise these operations. Although there exist previous contributions that either address filtering performance or identification of abnormal sensory observations, they do not provide a complete treatment of both problems at once. In this work we present a statistical approach that allows us to study and quantify the impact of abnormal observations from range sensors on the performance of Bayesian filters. For that, we formulate the estimation problem from a generic perspective (abstracting from concrete implementations), analyse the main limitations of common robotics range sensors, and define the factors that potentially affect the filtering performance. Rigorous statistical methods are then applied to a set of simulated experiments devised to reproduce a diversity of situations. The obtained results, which we also validate in a real environment, provide novel and relevant conclusions on the effect of abnormal range observations in these filters.

Perception models for selective harvesting robots in fruit and vegetable production

“…Commonly used probabilistic models for robot localisation are the recursive Bayesian filters (Castellano-Quero et al, 2020). These filters can be used to approximate the state of the robot (which can be its position and orientation) through a process of prediction and correction.…”