Opportunities of Self Supervised Learning for GNSS: Evaluation of a Deep Learning-Enhanced PVT Algorithm
This paper proposes a deep learning-enhanced PVT algorithm that combines supervised range correction prediction with JEPA-based self-supervised pretraining to effectively mitigate multipath interference and improve positioning accuracy in dense urban environments using unlabelled GNSS data.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the dense canyons of modern cities, finding one's way with a satellite signal is a game of echoes. Global Navigation Satellite Systems, or GNSS, rely on signals traveling in a straight line from space to a receiver on the ground. However, when a receiver is surrounded by tall buildings, these signals often bounce off glass and concrete before arriving. This creates a delayed copy of the original signal, a phenomenon known as multipath interference. The receiver, confused by the arrival time of these echoes, calculates a position that is slightly off. In open country, this error is negligible, but in a crowded urban center, it can push a navigation system hundreds of meters away from the true location. For autonomous vehicles, delivery drones, or even a pedestrian trying to find a specific storefront, this inaccuracy is a critical barrier. The challenge has long been that the environment causing the error changes constantly; a building that blocks a signal today might be different tomorrow, and the specific way a signal bounces depends on the unique shape of the receiver's surroundings.
Researchers at Abbia GNSS Technologies in Toulouse have developed a new approach to solve this problem by teaching a computer to recognize the subtle fingerprints of these signal echoes. Instead of relying solely on pre-programmed physics models, which often fail to capture the complexity of a real city, they used a deep learning system to learn directly from the data. The core of their innovation is a method called self-supervised learning. In traditional machine learning, a computer learns by being shown thousands of examples with the correct answers already written down, such as a photo of a street paired with its exact location. Gathering this kind of "ground truth" data for driving is incredibly difficult and expensive, requiring high-end reference equipment and perfect conditions. The team realized they could bypass this bottleneck by first letting the computer study vast amounts of unlabelled data—recordings of signals from standard receivers with no known location. By asking the system to predict missing parts of these signal patterns, they forced it to build a deep, internal understanding of how signals behave in different environments, much like a student learning the rules of a language by reading thousands of books before ever taking a test.
The researchers then took this pre-trained system and fine-tuned it with a smaller set of data where the correct locations were known. The result was a new algorithm that acts as a smart filter for satellite signals. When the system receives a signal, it analyzes a sequence of data points, looking for the specific signatures that indicate a signal has bounced. It then predicts two things for each signal: a correction to fix the distance measurement and a measure of how much it trusts that prediction. This dual output is crucial; if the system is unsure, it lowers the weight of that signal in the final calculation, preventing a single bad echo from ruining the entire position fix. To test this, the team drove through diverse urban environments in Toulouse, Tokyo, and Hong Kong, covering everything from wide boulevards to narrow streets lined with trees that cast deep shadows. They compared their new method against standard navigation techniques and against a version of their own algorithm that was trained only on labelled data, skipping the initial self-supervised learning phase.
The findings were clear and significant. In familiar driving scenarios, the new method improved accuracy, but the true breakthrough appeared in the harshest, most unpredictable conditions. When the system faced urban environments it had never seen before, the version trained with self-supervised learning remained reliable, while the standard version struggled. The data showed that the self-supervised approach prevented the system from taking shortcuts. Without the initial pre-training, the computer tended to rely too heavily on the immediate, raw numbers it was given, which worked well in easy conditions but failed when those numbers were corrupted by severe interference. The self-supervised model, having learned the broader patterns of signal behavior, understood the context better. It learned to trust the spatial relationships between signals rather than just the raw numbers, allowing it to make coherent corrections even when the initial position estimate was poor.
The researchers also looked inside the "brain" of their system to see how it made these decisions. They found that the model naturally learned to focus its attention on the parts of the sky where signals were most likely to be bouncing off the ground or low buildings, rather than the clear sky above. This spatial awareness emerged without being explicitly told to look there; the system discovered it on its own while learning to predict the missing parts of the signal data. This suggests that the method is not just memorizing specific streets but is learning the fundamental physics of how signals interact with the city. The study concludes that by using the vast amount of unlabelled data available from standard receivers, it is possible to build navigation systems that generalize far better to new and difficult environments. While the work does not claim to have solved the problem of urban navigation entirely, it demonstrates a powerful path forward: using self-supervised learning to anchor artificial intelligence in the physical reality of signal propagation, creating systems that are robust enough to navigate the complex, echo-filled cities of the future.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.