RA-LLO: Robust Adaptive Legged-LiDAR Odometry with Gaussian Process Motion Prior over Error States

초록

We propose RA-LLO, a real-time odometry framework tailored for highly dynamic legged robots, combining inertial sensors, joint kinematics, foot-contact sensing, and LiDAR data into a unified probabilistic approach. At its core, the method uses an error state Kalman filter with a Gaussian process-based continuous-time motion model, precisely discretized for accurate state prediction. Foot contacts are handled smoothly through confidence-weighted sensor inputs from pneumatic force sensors, ensuring stable stance estimation without sudden corrections. LiDAR scans are deskewed using a cubic B-spline fitted to predicted motion, aligning data accurately in time. Experiments with a Unitree Go2 robot demonstrated significant improvements in vertical drift and overall odometry accuracy compared to existing methods, running efficiently with low-drift state estimation suitable for challenging, unstructured terrains.

키워드