23 May 2025 in IEEE Robotics and Automation Letters
Image-goal navigation is a critical task in autonomous visual navigation, requiring the robot to navigate to a target localization specified by an image. Previous works using data-driven methods achieve great success while they mostly leverage simple network architecture and train it from scratch, which limits the navigation performance in challenging situations, such as multiple turns or varying lighting conditions. In this letter, we thoroughly analyze the essential features for visual navigation and design an effective network to achieve optimal navigation performance. In particular, we leverage a pretrained foundation model for feature extraction, introduce cross attention for goal encoding and propose a token attention mechanism to dynamically assign weights to different tokens. The proposed model achieves excellent navigation performance in unseen environments. Experiments in real world demonstrate that our method achieves a success rate of 87%, 40% improvement over the state-of-the-art methods.
9 February 2026 in IEEE Transactions on Instrumentation and Measurement
Dead-reckoning (DR) algorithms using inertial measurement unit (IMU) and wheel encoders can continuously provide positioning information in scenarios where GNSS signals are lost, providing strong autonomy and playing a critical role in vehicle autonomous navigation. However, a limitation of this method is that errors accumulate over time, resulting in decreasing positioning accuracy and even significant drift during long-term use. This article proposes a learning-based DR algorithm that leverages an encoder–decoder neural network integrating a multimodal attention module (MMAM) and a coarse-to-fine fusion strategy. This approach effectively extracts and fuses motion characteristics from IMU and wheel encoders data, thereby accurately estimating the 3-D vehicle’s instant velocity in the world frame and quantifying its uncertainty. The predicted velocity is then used as the pseudo-measurement in an invariant extended Kalman filter (IEKF), where the uncertainty is directly mapped into a dynamical measurement noise covariance matrix that adjusts the filter update. Real-world experiments demonstrate that our method outperforms existing techniques, reducing average positioning error by approximately 76% compared to the traditional INS + ODO baseline and guaranteeing robust long-term localization in dynamic environments.