AI-Powered Inertial Sensing

We developed and applied artificial intelligence-powered inertial sensing algorithms, including both hybrid learning and end-to-end learning approaches, across various platforms and applications. Our algorithms significantly enhance common navigation and estimation tasks, opening new possibilities for accurate and robust navigation. They include autonomous underwater vehicle navigation for ocean exploration and protection, pure inertial navigation for quadrotors enabling accurate positioning in urban environments and indoors, pedestrian dead reckoning using a smartphone enabling the development of applications for safety and health, attitude and heading reference systems, land vehicle navigation, illness detection in dogs, and multi-platform algorithms such as denoising. The diversity in our AI-powered inertial sensing and achievements in this field puts us in the top inertial labs worldwide.  Our 5-year efforts resulted in 61 journal and conference publications in the AI-powered inertial sensing domain. All our publications include experimental validation using our lab equipment, which includes an autonomous boat (self-designed), mobile robots, quadrotors, and an autonomous underwater vehicle.

Navigating the Depth: Deep Learning for Autonomous Underwater  Navigation 

Oceans cover about two-thirds of the surface of the Earth and have a great impact on mankind, and will continue to do so in the future. Autonomous underwater vehicles (AUV) have shown significant advancements in recent years, leading to breakthroughs in ocean exploration, underwater structure inspection, marine research, and marine resources management. For any AUV mission to succeed, accurate and robust navigation is essential. Commonly, a Doppler velocity log (DVL) and inertial sensors are fused to meet this requirement. The inertial/DVL fusion aims to use the advantages of the two systems, overcoming their weaknesses. When combined, the two systems can provide a bounded navigation solution, including the position, velocity, and orientation of the AUV, over long periods.

At ANSFL, we have developed hybrid deep-learning solutions for several challenging AUV navigational tasks. The first is an adaptive Kalman-informed transformer combining the strengths of the well-established extended Kalman filter and leveraging well-known deep learning characteristics. Our approach introduces a novel Kalman-informed loss configuration designed to emulate filter principles, enhancing process noise covariance estimation accuracy. Secondly, to improve the estimation of the velocity vector and to cope with real-world partial or complete loss of DVL beam measurements, we offer end-to-end multi-head architectures designed for both DVL velocity vector estimation and strategies to fuse them in the AUV navigation filter. Next, a simple, yet efficient, end-to-end regression network for the DVL calibration process is developed. The network is a convolution-based framework that employs a two-dimensional dilated convolution kernel designed to process input data uniquely. Finally, we present an end-to-end deep learning approach to estimate the AUV acceleration vector based on past DVL velocity measurements. This measurement is later used to update the navigation filter and contributes to its stability and performance. To validate our proposed approach, we recorded inertial and DVL readings from our AUV during sea experiments. The dataset contains five hours of recordings made under different mission parameters and sea conditions. Our results demonstrate that our proposed learning approaches outperform model-based methods in terms of accuracy and robustness of AUV navigation. As a result of our data-driven approaches, we can enhance navigation accuracy and reliability in AUV applications, contributing to more efficient and effective underwater missions through improved accuracy and reliability. Currently, we continue to explore exciting new research in those fields.

Neural Dead- Reckoning and Fusion

Interest in autonomous platforms such as quadrotors and mobile robots has increased significantly in recent years. One of the reasons is the ability to navigate accurately using low-cost sensors. Such autonomous platforms are used for a variety of applications both outdoors and indoors. They include deliveries, surveillance, transportation, mapping, and more. To accomplish these tasks, an accurate navigation solution is required. To that end, inertial sensors are fused with a global navigation satellite system signals (outdoors) or vision (indoors/outdoors). Due to environmental or sensor constraints, the navigation solution may rely only on inertial sensors.  For example, GNSS outages or poor lighting conditions. As a consequence, the navigation solution drifts in time.

To cope with such situations we propose a framework for quadrotor navigation based only on inertial sensors: quadrotor dead reckoning (QDR). Motivated by the pedestrian dead reckoning approach (PDR), basic concepts are adopted for the QDR framework. To that end, in situations of pure inertial navigation instead of moving in a straight line, the quadrotor is flown in periodic trajectories (PTSs) to emulate a walking pedestrian. In that manner, similar to step-length detection and estimation in PDR, the quadrotor peak to peak distance is estimated, leading to an accurate navigation solution compared to a pure inertial one. Our goal in this research was to use only low-cost inertial sensors, supported by appropriate techniques and algorithms, to enhance quadrotors' pure inertial navigation.

With periodic dynamics, inertial deep learning approaches can capture motion more effectively and provide accurate dead reckoning for drones and mobile robots. Thus, in a set of papers we propose approaches to deep-learning assisted dead reckoning to obtain accurate platform positioning. Later, we demonstrated how neural-position updates can be used as an external update to the navigation filter. We further showed that this update is beneficial even in situations where GNSS is available. Active research initiatives are dedicated to further explore emerging research frontiers within these disciplines.