Abstract
We present a method for parameter learning in a Gaussian variational inference setting using only noisy measurements, without requiring groundtruth data. This approach is demonstrated within the context of vehicle trajectory estimation by extending the Exactly Sparse Gaussian Variational Inference (ESGVI) framework. Our contribution involves learning the covariances for motion and sensor models, specifically utilizing a white-noise-on-acceleration motion model and an Inverse-Wishart prior for measurement covariances. Tested on a 36 km lidar dataset and standard pose graph optimization benchmarks, the method achieves high-quality state estimates and demonstrates significant robustness to outliers and false loop closures.
Introduction
Probabilistic state estimation is fundamental to robot navigation, yet it relies heavily on accurately tuned noise models. Traditionally, these parameters are either hand-tuned or learned using expensive groundtruth data. This paper addresses the "chicken-and-egg" problem of estimating a robot's path while simultaneously learning the noise characteristics of its sensors and actuators. By extending the ESGVI framework, we enable the system to learn these parameters directly from noisy observations, making the process more adaptable to new platforms and resilient to real-world data imperfections.
Methodology: ESGVI and Parameter Learning
The proposed framework combines Exactly Sparse Gaussian Variational Inference (ESGVI) with an Expectation-Maximization (EM) loop. In the E-step, the algorithm estimates the most likely robot trajectory (latent state) given fixed noise parameters. In the M-step, the trajectory is held fixed while the noise parameters—specifically the covariances—are optimized to better fit the observed data.
A key innovation is the use of an Inverse-Wishart (IW) prior for sensor noise. Rather than assuming a static covariance, the system treats the covariance as a random variable. This allows for automatic outlier rejection; if a measurement is significantly inconsistent with the estimated path, the system reduces its influence. To prevent the optimization from collapsing (uncertainty reaching zero), the authors implement a volume constraint on the determinant of the scale matrix.
Experimental Setup and Results
The method was validated on a 36 km lidar dataset (Route A for training, Route B for testing) and the Bicocca 25b dataset for pose graph optimization.
Performance and Robustness
- Groundtruth Independence: Training without groundtruth yielded a translational error of 0.2335m, performing slightly better than training with incomplete groundtruth (0.2368m).
- Outlier Mastery: In scenarios with 5% artificial outliers, the proposed IW-prior method maintained an error of ~0.23m, whereas standard static models failed completely (errors > 5m).
- SLAM Stability: The framework successfully resolved pose graphs containing numerous false loop closures that typically cause divergence in standard SLAM systems.
Discussion and Future Work
The results demonstrate that the internal consistency of a factor graph is sufficient to learn noise parameters without a perfect reference. While the system currently requires manual setting of two metaparameters (degrees-of-freedom and the determinant constraint), future work aims to automate this process. Furthermore, the authors suggest this framework could be expanded to learn complex robot models represented by Deep Neural Networks (DNNs), bridging the gap between classical estimation and modern deep learning.
Summary updated