Heteroscedastic Gaussian Process Regression for Modeling Range Sensors in Mobile Robotics

Abstract

In probabilistic approaches to mobile robot navigation, the development of measurement models plays a crucial role as it directly influences the efficiency and the robustness of the robot’s performance in a great variety of tasks including localization, tracking, and map building [1]. Among the most popular types of sensors used are range finders, which measure distances to nearby obstacles relative to certain (possibly multivariate) bearing angles. Probabilistic measurement models for this kind of sensor, such as beam models (aka. ray-casting models) and likelihood fields (aka. end point models), typically assume independency between individual range measurements. This leads to a series of practical limitations such as overly peaked observation likelihood functions for high density range scans or degrading performance in highly cluttered environments. To overcome this, we propose a novel probabilistic measurement model for range finders, called Gaussian Beam Processes. Gaussian Beam Processes treat the measurement modeling task as a nonparametric Bayesian regression problem and solve it using Gaussian processes [2]. The major advantage of this approach lies in the smoothness of the model, resulting from the representation of correlations between adjacent beams using covariance functions. Moreover, the Gaussian process treatment results in a sound probabilistic measurement model with a pool of well-established techniques for likelihood estimation and range prediction for an arbitrary number of beams. Standard Gaussian processes, however, assume a constant noise term over the domain. For modeling range sensor measurements, the variance of range values in each beam direction is, along with its mean value, an important feature of the sought-after distribution of range measurements. Inspired by Goldberg et al. [3], we therefore extended the standard Gaussian process framework to deal with heteroscedasticity, i.e., non-constant noise. Like Goldberg et al., we model the log noise variances explicitly using a second Gaussian process. In contrast to their work, however, we do not use a time consuming Markov chain Monte Carlo method to estimate the posterior noise variances but a fast most-likely-noise approach. We implemented the proposed system and evaluated it using a real robot as well as simulation for one of the classical tasks of robotics, namely mobile robot localization. The experimental results using a particle filter for mobile robot localization demonstrate the effectiveness of Gaussian Beam Processes in comparison to previous approaches.