Artificial Intelligence making its way into PNT

Last Updated: 12/09/2024 14:12     Created at: 12/09/2024 14:12

Article 1: "Generative AI Reshaping the PNT Landscape"

The 1980s may have been the dawn of AI research, but recent breakthroughs have ignited a quantum leap in its capabilities. One key innovation, the transformer architecture, has dramatically improved neural network efficiency. Imagine a network that can focus its attention, just like you do when reading a complex text - that's the power of the "self-attention mechanism" introduced in the paper, "Attention is All You Need". This, coupled with the decreasing cost of computing, can now be achieved at a fraction of the price, making advanced AI both accessible and practical. But that's not all.  The plummeting cost of computing has turned the tide. We've seen this in the rapid development of large language models (LLMs) like ChatGPT (from OpenAI), Gemini (from Google), Claude AI (from Anthropic), LLaMA (from Meta), and xAI/Grok (an initiative by Elon Musk). These chatbots showcase the impressive progress in AI for natural language processing, and it's a testament to the broader field's potential to revolutionize numerous industries. Similarly, generative AI, a subfield of AI, empowers algorithms to analyse vast datasets, identify intricate patterns, and create entirely new outputs. This opens the door for the PNT industry to harness the power of advanced AI – but how exactly will it transform the way we approach Positioning, Navigation, and Timing? 

This series of three articles will delve into the exciting world of AI-powered PNT. Part 1 focuses on generative AI, the underlying technology behind many PNT-related AI solutions. Parts 2 and 3 will explore other emerging AI techniques and their practical applications within PNT.

Before we dive deeper into the applications of generative AI in PNT, it's essential to understand a key concept: the self-attention mechanism. This mechanism is a cornerstone of transformer architectures, which have revolutionized natural language processing and, subsequently, generative AI. In essence, self-attention allows a neural network to weigh the importance of different parts of an input sequence when processing it. For example, in a sentence, the attention mechanism helps the network focus on relevant words and their relationships to understand the overall meaning.

Generative AI, powered by the self-attention mechanism, excels at processing multi-source PNT data, including GNSS signals, inertial measurements, terrestrial beacons, and crowdsourced information. In challenging environments like urban canyons, it can mitigate the effects of signal multipath and Non-Line-of-Sight (NLOS) conditions, significantly enhancing positioning accuracy. This technology can even generate synthetic GNSS constellations to test receiver performance under diverse scenarios or create high-fidelity simulations of complex radio frequency (RF) environments. As seen in the NAVISP project EL1-035 Machine Learning to model GNSS systems – EGNOS Performance Prediction System, machine learning techniques can be used to model GNSS system segments, enabling the creation of more realistic and representative simulations.

Beyond enhanced accuracy, generative AI excels at predictive modelling, such as for real-time traffic forecasting and route optimization in navigation applications.  These models account for factors like signal obstruction, atmospheric conditions, and historical congestion patterns.  Similarly, for timing applications, generative AI can predict and compensate for clock errors and time transfer uncertainties across distributed systems. Pushing the boundaries even further, generative AI finds applications in advanced mapping and localization.  It can generate detailed 3D maps from sparse point clouds, enhancing Simultaneous Localization and Mapping (SLAM) techniques for autonomous systems.  For professional surveying and geodesy, generative AI offers improvements in the accuracy of coordinate transformations and geoid models. In accessibility applications, it can create context-aware, spatially anchored audio guidance, vastly improving navigation for visually impaired users in complex environments. A  NAVISP project called NaviBlind, also integrated the use of AI to ensure reliable data for navigation by using machine learning to identify sidewalks, crosswalks and Audible Pedestrian Signals (APS). 

Indoor Navigation and Beyond

Indoor navigation, long a challenge due to the absence of GNSS signals, can be enhanced through generative AI by leveraging multimodal data sources, including building blueprints, sensor data, and user interactions. Generative AI models can create highly accurate and detailed indoor maps. These maps can be dynamically updated to reflect changes in building layout, occupancy, and environmental conditions. Moreover, generative AI can enhance the user experience by providing personalized navigation guidance. By analysing user preferences, behaviour patterns, and real-time data, AI algorithms can suggest optimal routes, considering factors such as crowd density, elevator wait times, and accessibility needs. 

KalmanNet: Combining Strengths 

As the PNT Industry and academia explore new frontiers, a new wave of AI architectures is emerging to bridge the gaps. One such innovation is the state-of-the-art neural network for PNT called KalmanNet, a neural network fusing Kalman filtering and deep learning. By combining the strengths of Kalman filtering with deep learning, KalmanNet enables adaptive and robust state estimation through data-driven optimization of filter parameters. When integrated with generative AI, KalmanNet can be further enhanced, creating a synergistic approach that holds the potential to significantly advance the field of PNT. Generative models can be used to create synthetic training data, improving the network's ability to handle complex and rare scenarios. Additionally, generative AI can assist in model selection and hyperparameter tuning for KalmanNet, accelerating development and improving performance. This approach has shown promising results in various PNT applications, including GNSS signal processing, inertial navigation, and sensor fusion etc. 

The Road Ahead

Generative AI is undeniably on an upward trajectory, currently residing in the "Peak of Inflated Expectations" on the Gartner hype cycle. While this translates to high-profile successes and significant investments, it's crucial to navigate this phase with a focus on practical applications within PNT. By understanding the current stage of technology adoption, harnessing the true potential of generative AI requires addressing key challenges:

• Data Quality: Training generative models requires high-quality, diverse PNT datasets (raw sensor data, processed sensor data round truth data, environmental data etc.) Collaborative efforts across the industry are essential to address potential data limitations.
• Bias Mitigation: Generative models can inherit biases present in training data. Techniques for bias detection and mitigation are necessary to ensure fair and reliable AI-enhanced PNT solutions.
• Explainability/Verifiability: Understanding how generative models arrive at their outputs is critical for building trust and ensuring the robustness of AI-enhanced PNT systems. Advancements in explainable AI (XAI) are crucial in this regard. Verifiability on the other hand involves rigorous validation of AI model performance within complex and evolving operational domains. Formal verification techniques are essential tools for scrutinizing AI model behaviour against predefined criteria.

By addressing these challenges and focusing on practical applications like synthetic data generation, multi-sensor data fusion optimization, and personalized navigation, PNT industry can usher a new era of robust, adaptable, and user-centric navigation thanks to generative AI.