Vehicle navigation has long relied on satellite-based systems like Global Navigation Satellite System (GNSS), yet these signals often falter in covered areas such as tunnels, underground garages, and urban canyons. While high-end vehicles use a suite of sophisticated sensors to fill this gap, smartphones depend on low-cost inertial sensors, which suffer from significant drift and inaccuracies over time. With the rise of deep learning, researchers have begun using artificial intelligence to infer movement directly from noisy Inertial Measurement Units (IMU) data. Yet, achieving both accuracy and robustness with such limited input remains a formidable task. Due to these challenges, there is an urgent need for a self-contained, AI-enhanced navigation solution that works even when satellites can’t.
A collaborative team from Wuhan University and Chongqing University has unveiled a smartphone-only inertial navigation framework published (DOI: 10.1186/s43020-025-00168-7) in Satellite Navigation in June 2025. Their approach, dubbed DMDVDR (Data- and Model-Driven Vehicle Dead Reckoning), uses a custom-designed deep neural network—AVNet—to extract motion cues from inertial sensor data and integrates them into an invariant Kalman filter for accurate trajectory estimation. The system operates without any Global Positioning System (GPS) input, making it uniquely suited for GNSS-denied environments such as underground tunnels or parking facilities.
At the heart of this new framework lies AVNet, a hybrid deep learning architecture combining convolutional and recurrent layers. AVNet processes raw data from a smartphone’s IMU to estimate real-time vehicle orientation and velocity. These pseudo-measurements are then fused into a mathematical filter known as the Invariant Extended Kalman Filter (InEKF), which compensates for sensor noise and drift by integrating both model-based and AI-inferred data.
To further refine the system’s performance, the researchers introduced a data-driven filter parameter adapter that dynamically learns optimal noise profiles, allowing the system to adapt to various driving conditions. Tested in a parking lot using consumer smartphones, the method outperformed existing solutions by a wide margin, achieving a horizontal translation error of just 0.4%. Even in more complex scenarios—such as reverse parking or repeated turns—the system maintained stability and accuracy. Beyond local experiments, the researchers validated their system on real-world tunnel data from the Google Smartphone Decimeter Challenge. The results showed only 0.64% positional drift after 578 meters of GPS signal loss, a testament to the framework’s resilience. By merging AI with classical control theory, the study offers a compelling solution for robust vehicle localization when satellite signals fail.
“Our aim was to empower ordinary smartphones to deliver extraordinary navigation—even where GPS stops working,” said Dr. Ruizhi Chen, the study’s senior author. “By combining deep learning with proven filtering techniques, we’ve created a system that doesn’t just work in theory—it performs reliably in real-world conditions like tunnels and underground lots. This is a leap forward for AI-driven mobility using everyday consumer devices.”
The proposed DMDVDR framework could revolutionize smartphone-based navigation systems by extending their usability into GPS-deprived areas. Potential applications include autonomous parking assistance, fleet management in covered facilities, and safer navigation in tunnels or dense urban environments. Because the system runs solely on smartphone sensors, it offers a scalable and low-cost alternative to complex in-vehicle navigation hardware. Furthermore, its compatibility with GNSS and other sensors makes it an ideal candidate for hybrid indoor-outdoor localization platforms. As smart mobility continues to evolve, solutions like DMDVDR will play a crucial role in ensuring uninterrupted, precise, and intelligent navigation for both personal and commercial transportation.
###
References
DOI
10.1186/s43020-025-00168-7
Original Source URL
https://doi.org/10.1186/s43020-025-00168-7
Funding Information
This study was supported by the National Key Research and Development Program of China (grant nos. 2023YFB3906600), and the NSFC (grant no. 42201460).
About Satellite Navigation
Satellite Navigation (E-ISSN: 2662-1363; ISSN: 2662-9291) is the official journal of Aerospace Information Research Institute, Chinese Academy of Sciences. The journal aims to report innovative ideas, new results or progress on the theoretical techniques and applications of satellite navigation. The journal welcomes original articles, reviews and commentaries.