Learned IMU Bias Prediction for Invariant Visual Inertial Odometry
Abdullah Altawaitan, Jason Stanley, Sambaran Ghosal, Thai Duong, Nikolay Atanasov
AI summary
Problem
Including IMU bias in the Kalman filter state breaks the Lie group symmetry required for invariant filters, degrading convergence and robustness. Standard random-walk bias models also fail to capture the slow, time-varying drift characteristic of real inertial sensors.
Approach
A sequence-to-sequence neural network learns to predict gyroscope and accelerometer biases directly from a sliding window of raw IMU data. This external prediction corrects the measurements before they enter an invariant multi-state constraint Kalman filter, preserving the filter's mathematical symmetry.
Key results
- Decouples bias estimation from the filter state to preserve Lie group symmetry
- Achieves real-time, robust visual-inertial odometry on EuRoC and Aerodrome datasets
- Maintains accurate state estimation during extended visual feature loss
- Produces stable, physically consistent bias estimates superior to standard EKF baselines
Why it matters
Provides a reliable localization solution for autonomous robots operating in visually degraded or dynamic environments where traditional VIO systems fail.
Abstract
Autonomous mobile robots operating in novel envi- ronments depend critically on accurate state estimation, often utilizing visual and inertial measurements. Recent work has shown that an invariant formulation of the extended Kalman filter improves the convergence and robustness of visual-inertial odometry by utilizing the Lie group structure of a robot’s position, velocity, and orientation states. However, inertial sensors also require measurement bias estimation, yet introducing the bias in the filter state breaks the Lie group symmetry. In this paper, we design a neural network to predict the bias of an inertial measurement unit (IMU) from a sequence of previous IMU measurements. This allows us to use an invariant filter for visual inertial odometry, relying on the learned bias prediction rather than introducing the bias in the filter state. We demonstrate that an invariant multi-state constraint Kalman filter (MSCKF) with learned bias predictions achieves robust visual-inertial odometry in real experiments, even when visual information is unavailable for extended periods and the system needs to rely solely on IMU measurements.