Reproduce “A Plug-and-Play Learning-based IMU Bias Factor for Robust Visual-Inertial Odometry” on OpenResearch

About this paper

Yang Yi, Kunqing Wang, Jinpu Zhang, Zhen Tan, Xiangke Wang, Hui Shen, Dewen Hu

Researchers at the National University of Defense Technology introduce a learning-based Inertial Prior Network (IPNet) that infers a stable IMU bias prior from raw inertial data, enhancing Visual-Inertial Odometry (VIO) robustness. Integrating IPNet as a "plug-and-play" factor into VIO systems improved localization precision by 46% ATE-RMSE and 48% RPE-RMSE on public datasets and prevented trajectory divergence in visually challenging environments.

Key insights

• Platform-specific motion patterns, even from low-cost IMUs, implicitly contain information that can be exploited to infer underlying IMU biases.
• Providing a stable, non-recursive, per-sequence IMU bias prior can effectively constrain VIO optimization and prevent error propagation when visual information is ambiguous or unreliable.
• A learned bias prior can be effectively incorporated as a regularization term in a VIO optimization framework without requiring modifications to the core VIO architecture.

Summary by alphaXiv.