Publication · 2026
PLATO: Privileged Learning from Accurate Trajectory Observations for Neural Inertial Preintegration
H Li, Q Liu, Y Peng, K Miao, J Wu
2026 Conference on Robot Learning (CoRL), 2026
Abstract
Neural inertial odometry has demonstrated strong potential for motion estimation in challenging environments, yet inertial-only preintegration remains sensitive to IMU bias and uncertainty. PLATO leverages accurate trajectory observations during training to jointly learn IMU bias dynamics, modeled by a Neural ODE, and gyroscope and accelerometer noise covariances. A structure-aware optimization scheme combines forward differentiation with a tailored double-adjoint method, enabling memory-efficient likelihood optimization over nested bias dynamics and IMU preintegration rollouts. Experiments on EuRoC and an underwater robotic platform demonstrate improved inertial propagation and downstream odometry under visual degradation and intermittent sensing failures.