[Paper Review] A Spectral Learning Approach to Range-Only SLAM

Byron Boots, Geoff Gordon|arXiv (Cornell University)|Jul 10, 2012

Robotics and Sensor-Based Localization25 references20 citations

TL;DR

This paper proposes a spectral learning approach to range-only SLAM that formulates the problem as a low-rank matrix factorization, enabling a linear, non-iterative solution via singular value decomposition (SVD). The method avoids local optima and linearization errors common in EKF/EIF and MHT, achieving near-batch-optimization accuracy at orders-of-magnitude lower computational cost, with empirical results showing it outperforms EKF, FastSLAM, and ROP-EKF while being competitive with batch optimization when combined with refinement.

ABSTRACT

We present a novel spectral learning algorithm for simultaneous localization and mapping (SLAM) from range data with known correspondences. This algorithm is an instance of a general spectral system identification framework, from which it inherits several desirable properties, including statistical consistency and no local optima. Compared with popular batch optimization or multiple-hypothesis tracking (MHT) methods for range-only SLAM, our spectral approach offers guaranteed low computational requirements and good tracking performance. Compared with popular extended Kalman filter (EKF) or extended information filter (EIF) approaches, and many MHT ones, our approach does not need to linearize a transition or measurement model; such linearizations can cause severe errors in EKFs and EIFs, and to a lesser extent MHT, particularly for the highly non-Gaussian posteriors encountered in range-only SLAM. We provide a theoretical analysis of our method, including finite-sample error bounds. Finally, we demonstrate on a real-world robotic SLAM problem that our algorithm is not only theoretically justified, but works well in practice: in a comparison of multiple methods, the lowest errors come from a combination of our algorithm with batch optimization, but our method alone produces nearly as good a result at far lower computational cost.

Motivation & Objective

• To address the limitations of EKF, EIF, and MHT methods in range-only SLAM, which suffer from linearization errors and local optima due to highly nonlinear and non-Gaussian posteriors.
• To develop a statistically consistent, non-iterative SLAM algorithm that avoids local minima and does not require linearization of motion or measurement models.
• To provide a theoretically grounded, computationally efficient alternative to batch optimization, which is accurate but slow and prone to local optima.
• To demonstrate that spectral SLAM can serve as a strong initialization for batch optimization, combining theoretical guarantees with practical performance.

Proposed method

• Formulates range-only SLAM as a matrix factorization problem where observations are linearly related to a 4- or 7-dimensional state space, enabling spectral learning.
• Uses a fixed-rank singular value decomposition (SVD) to estimate the state space and factorization, leveraging the spectral system identification framework for statistical consistency and no local optima.
• Constructs a data matrix from history and future features of range measurements, with the state as a bottleneck that minimizes prediction error under a rank constraint.
• Applies a rank-7 SVD to the observation matrix to recover robot poses and landmark positions, with automatic correction for odometry bias and sensor height offsets via state space expansion.
• Integrates the method with batch optimization by using spectral SLAM results as initialization, improving convergence and solution quality.
• Employs a feature construction method to interpolate missing range readings, enabling application to sparse datasets.

Experimental results

Research questions

• RQ1 Can a spectral learning approach achieve better accuracy and computational efficiency than EKF, EIF, and MHT in range-only SLAM without relying on linearization?
• RQ2 Does a matrix factorization formulation of SLAM lead to a method that is free of local optima and statistically consistent?
• RQ3 Can spectral SLAM serve as an effective initialization for nonlinear batch optimization, improving convergence and solution quality?
• RQ4 How does the method perform on real-world data with sparse, noisy range measurements and odometry drift?

Key findings

• Spectral SLAM achieved localization RMSE of 0.79m on the full path in Plaza 1 and 0.35m in Plaza 2, comparable to batch optimization (0.79m and 0.33m respectively), despite being orders of magnitude faster.
• The method outperformed Cartesian EKF (0.94m), FastSLAM (0.73m), and ROP-EKF (0.65m) on the last 10% of the path in Plaza 1, with lower full-path error than all EKF-based methods.
• When combined with batch optimization, spectral SLAM produced the best results—0.54m RMSE in the best 10% of Plaza 2, below the best batch-only result of 0.20m—indicating strong initialization quality.
• Runtime for spectral SLAM was approximately 3–4 orders of magnitude faster than Gauss-Newton batch optimization, taking ~100 seconds on Plaza 1 versus ~2.5 hours for batch.
• The algorithm demonstrated robustness to odometry bias and sensor height offsets by automatically learning corrections through state space expansion.
• The method achieved high accuracy even on sparse datasets with up to 500 time steps between range readings, thanks to feature interpolation and SVD-based factorization.