MAC-I2: learned metrics-aware covariance for robust visual-inertial fusion

zhenjun_zhao · x · 2026-09-11

A robotics paper from CMU (Wenshan Wang, Sebastian Scherer et al.) presents MAC-I², which replaces predefined uncertainties in visual-inertial fusion with learned metrics-aware covariances that faithfully reflect actual measurement noise, letting vision and IMU compete on their own merits under illumination changes, dynamic objects and textureless scenes. Visually, learned feature-matching uncertainties are propagated into pose covariances; on the inertial side, an IMU model with a learnable initial covariance plus fine-tuning on a held-out subset handles early-stage error accumulation, supporting robust initialization and calibration.

Original post →

More from Embodied

Embodied channel →