Learned IMU Bias Prediction for Invariant Visual Inertial Odometry (2505.06748v1)

Abstract: Autonomous mobile robots operating in novel environments 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.

• The paper proposes an invariant Extended Kalman Filter (EKF) framework for Visual Inertial Odometry (VIO) that leverages learned IMU bias predictions without including bias in the filter state.
• A neural network model is used to predict IMU biases from historical measurements, providing high-frequency corrections independent of visual data.
• Experimental validation on datasets like EuRoC and Aerodrome demonstrates performance comparable to or superior to existing VIO methods, especially in scenarios with degraded visual information.

Overview of Learned IMU Bias Prediction for Invariant Visual Inertial Odometry

The paper entitled "Learned IMU Bias Prediction for Invariant Visual Inertial Odometry" addresses the complex problem of improving visual-inertial odometry (VIO) by predicting inertial measurement unit (IMU) biases using a learned model. The work expands upon existing methods by integrating a neural network to estimate bias from historical IMU data, which enhances the state estimation accuracy of mobile robots, especially in scenarios with degraded visual information.

Key Contributions

1. Invariant EKF Framework: While traditional VIO methods include IMU bias in the filter state, potentially impacting the Lie group symmetry, this paper introduces a novel approach that leverages an invariant filter by excluding the bias from the filter state. The proposed invariant EKF operates effectively by incorporating learned bias predictions, thereby maintaining the symmetry properties desirable for robust state estimation.
2. Neural Network for Bias Prediction: The paper employs a neural network architecture to predict IMU biases based on previous IMU measurements. This approach harnesses sequence-to-sequence modeling to provide high-frequency bias corrections without reliance on visual inputs during difficult periods.
3. Empirical Validation: The research demonstrates through a series of experiments on the EuRoC and Aerodrome datasets that their method achieves comparable or sometimes superior performance to established VIO techniques like MSCKF and MSCEqF. Importantly, it shows significant robustness in scenarios where visual input is temporarily unavailable, underscoring the flexibility of the proposed method in maintaining reliable odometry.

Technical Insight

The invariant EKF framework used in this work represents an advancement over traditional EKF-based VIO methods by decoupling bias estimation from the filter state while still leveraging the theoretical advantages of the Lie group properties. This approach ensures that the IMU-based state propagation remains robust and independent of visual feature extraction failures.

The neural network employed for bias prediction includes architectures such as ResNet and TCN, which are capable of capturing the sequential nature of IMU data effectively. The selection of ResNet as the primary architecture is justified through empirical evaluation, signifying its capability to model the slow time-varying nature of the IMU biases more effectively compared to other configurations.

Numerical Results and Implications

The numerical results provided in the paper illustrate the capabilities of the proposed method in terms of both position and orientation accuracy under scenarios with intermittent visual data. The method outperforms the conventional MSCKF, especially in environments with prolonged periods of visual degradation. This improvement is attributed to the sequence-to-sequence model's ability to learn bias corrections effectively, facilitating more accurate inertial integration and error correction.

The strong performance of this method in challenging conditions underscores its potential practical applications, particularly in autonomous operations where visual sensors might frequently encounter occlusions, lighting changes, or motion blur.

Future Developments

The paper points toward future directions, such as incorporating measurement uncertainties into the learned model to further enhance robustness. This exploration could lead to more adaptive VIO systems that dynamically adjust to varying sensor reliability, improving estimations further under unpredictable conditions.

Moreover, incorporating additional inputs, such as image features or past visual data, in the neural network model could extend the applicability of this method to a broader set of operational scenarios, potentially offering further improvements in scenarios with complex dynamic lighting and motion conditions.

In conclusion, the paper presents a sophisticated approach to enhancing VIO systems through learned IMU bias prediction while maintaining invariant Lie group properties in the state propagation, representing a substantial step forward in the domain of autonomous robotic navigation.