Post-Doctoral Research Visit – Learning crowd dynamics from real-world data

Type de contrat : CDD Contexte et atouts du poste The VirtUs team at the Inria Centre at the University of Rennes is internationally recognized for its work in crowd simulation and the study of collective human behaviour. This postdoctoral position is part of the FOUL-X project (Programme Inria Quadrant), which aims to develop a new generation of crowd simulators capable of automatically adapting to the specific dynamics of a given environment or situation. Current crowd simulation models rely on simplified, universal rules that fail to capture the diversity of behaviours observed in real-world settings. FOUL-X challenges this paradigm by exploring data-driven approaches that learn crowd dynamics directly from field observations. This requires addressing open scientific questions on how to represent crowd data, which learning architectures are best suited to capture collective behaviours, and how to evaluate the realism of learned simulations. This postdoc focuses on the development of machine learning models for crowd dynamics, working in close collaboration with the data acquisition activities of the project. The work will span the full pipeline from data representation to model learning and evaluation, with the ultimate goal of demonstrating an adaptive crowd simulator built on real-world data. Mission confiée Assignments: With the help of the VirtUs team and under the supervision of Julien Pettré, the recruited person will be tasked with developing machine learning approaches capable of automatically modelling crowd dynamics from real-world field data. The central objective is to demonstrate that a learning-based model can capture the variety of crowd dynamics observed across different sites and situations — a challenge that remains largely unexplored in the field. The expected outcome is a new class of crowd simulation models that can automatically adapt to a specific crowd dynamic, as opposed to the universal, simplified rules used by current simulators. For a better knowledge of the proposed research subject: A state of the art, bibliography and scientific references are available on the VirtUs team website: https://www.inria.fr/en/virtus Collaboration: The recruited person will work in close connection with the first postdoctoral researcher of the FOUL-X project, who is responsible for building the field dataset that will serve as the primary input for the modelling work. The postdoc will also interact regularly with a PhD student of the team developing the pedestrian tracking pipeline, whose outputs feed directly into the learning process. This close collaboration ensures that modelling choices are informed by the nature and constraints of the available data, and reciprocally, that data acquisition is guided by the requirements of the learning approaches. Responsibilities: The person recruited is responsible for the design, implementation and evaluation of machine learning models for crowd dynamics, working with the dataset progressively built during the project. The recruited person will take initiatives in exploring a range of modelling paradigms — including generative models, imitation learning, or physics-informed approaches — and will contribute to defining evaluation metrics adapted to the specific challenge of assessing the diversity of learned crowd dynamics. Steering/Management: The person recruited will be in charge of the modelling and learning activities of the FOUL-X project, from the initial design of data representations and learning architectures to the evaluation and dissemination of results at major scientific venues. Principales activités Phase 1 — Architecture design and preliminary learning (months 1–6) • Conduct a targeted review of existing approaches for data-driven crowd dynamics modelling, covering trajectory prediction, generative models, imitation learning, and physics-informed approaches • Define crowd data representations suited to machine learning, combining individual (positions, velocities), collective (density, flow), and environmental (obstacles, spatial layout) information • Select and implement the most promising learning architecture for crowd dynamics modelling, based on pre-existing datasets available in the team • Validate the technical functioning of the learning pipeline and establish baseline performance metrics Phase 2 — Learning diverse crowd dynamics from FOUL-X data (months 7–24) • Develop and iteratively refine machine learning models for crowd dynamics using the dataset progressively built by PDoc 1 across multiple acquisition sites • Address the challenges of learning from limited and partially observable real-world data, exploring techniques such as transfer learning, data augmentation, and weak supervision • Demonstrate the capacity of the models to capture and distinguish diverse crowd dynamics, as observed across different sites, populations, and spatial configurations • Contribute to the definition of evaluation metrics adapted to the ass