An Adaptive Graduated Nonconvexity Loss Function for Robust Nonlinear Least Squares Solutions

Kyungmin Jung,Thomas Hitchcox,James Richard Forbes

DOI: https://doi.org/10.1109/TRO.2024.3434169

2024-07-31

Abstract:Many problems in robotics, such as estimating the state from noisy sensor data or aligning two point clouds, can be posed and solved as least-squares problems. Unfortunately, vanilla nonminimal solvers for least-squares problems are notoriously sensitive to outliers. As such, various robust loss functions have been proposed to reduce the sensitivity to outliers. Examples of loss functions include pseudo-Huber, Cauchy, and Geman-McClure. Recently, these loss functions have been generalized into a single loss function that enables the best loss function to be found adaptively based on the distribution of the residuals. However, even with the generalized robust loss function, most nonminimal solvers can only be solved locally given a prior state estimate due to the nonconvexity of the problem. The first contribution of this paper is to combine graduated nonconvexity (GNC) with the generalized robust loss function to solve least-squares problems without a prior state estimate and without the need to specify a loss function. Moreover, existing loss functions, including the generalized loss function, are based on Gaussian-like distribution. However, residuals are often defined as the squared norm of a multivariate error and distributed in a Chi-like fashion. The second contribution of this paper is to apply a norm-aware adaptive robust loss function within a GNC framework. The proposed approach enables a GNC formulation of a generalized loss function such that GNC can be readily applied to a wider family of loss functions. Furthermore, simulations and experiments demonstrate that the proposed method is more robust compared to non-GNC counterparts, and yields faster convergence times compared to other GNC formulations.

Robotics

What problem does this paper attempt to address?

The paper primarily aims to address two key issues in the field of robotics related to the nonlinear least squares problem: the impact of outliers and the sensitivity to initialization errors.

The paper first points out the poor performance of traditional non-reduced least squares solvers when faced with outliers caused by sensor noise, data association errors, or model mis-specification. To mitigate the impact of outliers, robust loss functions are typically used, but these functions require pre-selection and tuning, and are sensitive to initial estimates. Recent research has proposed "Graduated Nonconvexity" (GNC), a method for handling large initialization errors with a specific choice of robust loss function.

The main contribution of the paper is the combination of adaptive kernels with the GNC optimization scheme, proposing a new robust least squares solving method. This method does not require manual selection and tuning of the loss function and is robust to both outliers and initialization errors. Specifically, the method uses an adaptive loss function and gradually increases nonconvexity through GNC to solve the nonlinear least squares problem.

Additionally, the paper extends this method by considering the gap between the actual pattern of residual distribution and the assumed one. By identifying the pattern of the residual distribution and applying the adaptive loss function to the residuals after this pattern shift, the effectiveness of the method is further improved.

Finally, through experimental validation on problems such as point cloud alignment, mesh registration, and pose graph optimization, it is demonstrated that the proposed GNC-ADAPT and GNC-AMB methods significantly improve robustness and performance compared to other non-GNC adaptive loss functions and GNC methods with specific loss functions.