진학프로석·박사 채용 접수 플랫폼

Université Gustave Eiffel is looking for a candidate to apply for a Postdoctoral Fellowship in the framework of the Marie-Sklodowska Curie Programme 2025.

The Candidate and Université Gustave Eiffel's supervisor will apply together to develop the following research project : Smart and Affordable Labeling for Multisensory Navigation Data Using Reinforcement Learning

1. Summary

The complexity of positioning scenarios and the requirements for location-based applications have significantly increased due to technological advancements, urbanization, and the broadening of use cases across industries. Applications in fields such as autonomous vehicles, indoor navigation, healthcare, augmented reality, and logistics all require increasingly sophisticated solutions to deliver precise, reliable, and adaptive experiences that account for dynamic and diverse conditions. At the same time, vast amounts of navigation data are generated from different navigation sensors, necessitating advanced methods for processing, and analyzing.

That is why, data-driven approaches are becoming increasingly popular in the localization and navigation domain due to their ability to address complex challenges that traditional methods struggle to model accurately. For instance, GNSS multipath and Non-Line-of-Sight (NLOS) reception errors are difficult to predict with conventional models, leading many researchers to explore Artificial Intelligence (AI)-based techniques for detecting and excluding faulty GNSS measurements [1-2]. AI models, particularly those leveraging machine learning, offer a more adaptive and dynamic approach to identifying and mitigating these errors, significantly enhancing the reliability of GNSS-based positioning systems. Additionally, parameters for navigation filters, such as the Extended Kalman Filter (EKF) and Factor Graph Optimization (FGO), are crucial for determining positioning performance, as they are highly sensitive to contextual factors like environmental conditions and device quality. Fine-tuning these filtering parameters can greatly improve the stability and robustness of location accuracy, and deep learning techniques show promise in automating and optimizing this process, making it more efficient and precise [3]. In the navigation and positioning domain, supervised learning techniques are often applied to ensure more controllable, predictable results, and better overall performance, providing an effective means to adapt to diverse conditions and system requirements.

The labels required for supervised learning, particularly high-quality labels, are often costly to obtain due to factors such as the need for human labor, expensive equipment, or access to detailed map data. For example, labeling GNSS Line-of-Sight (LOS) and NLOS data typically relies on cumbersome devices like fish-eye cameras or 3D map-aided systems, which can introduce errors such as image segmentation inaccuracies, camera calibration errors, and map discrepancies. The main research question and objective of this research project is to propose a generalizable data labeling methodology that uses reinforcement learning (RL) to automatically generate high-quality labels. By utilizing only the reference trajectory, this approach aims to label intermediate quantities in particular sensor data quality and filtering parameters in a cost-effective and scalable way.

Reinforcement learning is particularly suited for this task because it enables an adaptive, autonomous learning process where the system can continuously refine its labeling strategy based on feedback from the environment. The main tasks in this research project include: 1) A thorough state-of-the-art on reinforcement learning focusing on its applications in localization and navigation systems, as well as its potential to address challenges in data labeling. 2) Modeling the navigation data labeling problem within the framework of reinforcement learning by carefully defining the core elements of the problem, including the state, actions, environments, and rewards. This leads to an appropriate methodology that allows reinforcement learning algorithms to autonomously optimize the labeling process, relying only on the reference trajectory, which is anyway needed for performance evaluation. In this way, the cost and efficiency of data labeling will be significantly reduced. 3) Two particular study will be conducted to demonstrate the effectiveness of the proposed methodology: a) sensor quality labeling, eg., GNSS and Inertial Navigation System (INS) data; b) filtering parameter labeling, e.g., EKF and/or FGO, ensuring that the filter remains robust and accurate under varying conditions

[1] Zhu, Ni, Chaimae Belemoualem, and Valérie Renaudin. "A Portfolio of Machine Learning-Based GNSS LOS/NLOS Classification in Urban Environments." In 2023 IEEE SENSORS, pp. 1-4. IEEE, 2023.
[2] García Crespillo, O., Ruiz-Sicilia, J. C., Kliman, A., & Marais, J. (2023). Robust design of a machine learning-based GNSS NLOS detector with multi-frequency features. Frontiers in Robotics and AI, 10, 1171255.
[3] Li, Shuo, et al. "Exploring the Potential of Deep Learning Aided Kalman Filter for GNSS/INS Integration: A Study on 2D Simulation Datasets." IEEE Transactions on Aerospace and Electronic Systems (2023).
[4] Zhu, N., Bouronopoulos, A., Leduc, T., Servières, M., & Renaudin, V. (2023, April). Evaluation of the Human Body Mask Effects on GNSS Wearable Devices for Outdoor Pedestrian Navigation Using Fisheye Sky Views. In 2023 IEEE/ION Position, Location and Navigation Symposium (PLANS) (pp. 841-850). IEEE.

2. Planned secondments

Secondments will be decided with the candidate

3. Planned duration of the project

24 months