Hybrid Multi-Camera Visual-Inertial Odometry for Robust Humanoid Navigation in Complex Environments

Visual-inertial odometry (VIO) remains essential for autonomous navigation, yet faces amplified challenges in humanoid robots due to their dynamic locomotion and complex environments. We propose a hybrid multi-camera VIO system based on Multi-State Constraint Kalman Filter (MSC-KF), targeting humanoid platforms. Our architecture fuses multi-camera and inertial data via MSC-KF to address bipedal motion dynamics and visual degradation, where a learning-based frontend ensures robust feature matching under motion blur and illumination changes and a Savitzky-Golay filter suppresses IMU noise during aggressive motions. Evaluations in an office environment with a long corridor demonstrate superior robustness against high dynamics, lighting variations, and texture-deficient regions. Comparative tests reveal stable tracking particularly in the long corridor section where conventional methods fail due to geometric repetition and sparse features, confirming practical advantages for legged autonomy.