Natural Gradient Gaussian Approximation Filter on Lie Groups for Robot State Estimation

A research team from Tsinghua University has introduced a novel state estimation algorithm called the Natural Gradient Gaussian Approximation on Lie Groups (NANO-L) filter. The core innovation addresses a fundamental challenge in robotics: accurately tracking a robot's position and orientation (its 'state') when it moves on curved mathematical spaces known as Lie groups, which is essential for systems like legged robots. Traditional filters rely on local linear approximations that accumulate error. NANO-L reformulates the entire filtering problem as a direct parameter optimization over a Gaussian variable, completely bypassing the need for linearization. This variable is mapped to the robot's pose via an exponential operator, providing a more geometrically consistent update.

The team further enhanced this approach with a natural gradient optimization scheme. This method uses the Fisher information matrix to account for the curvature of the underlying space, leading to more efficient and accurate updates. Crucially, for a common class of observation models used in robot localization, the researchers proved that NANO-L's covariance update has an exact, closed-form solution. This eliminates the need for iterative calculations in many practical scenarios, preserving the accuracy gains without a computational penalty.

Hardware validation on a commercial Unitree GO2 quadruped robot operating across different terrains demonstrated the filter's significant practical advantage. The NANO-L filter achieved approximately 40% lower estimation error compared to commonly used state-of-the-art filters, all while maintaining a comparable computational cost. This breakthrough in estimation fidelity is a key prerequisite for enabling more dynamic, reliable, and agile autonomous motion in complex real-world environments.