Projet ciblé HAMMER - PEPR Robotique - porté par G. Allibert et B. Mavkov

Le Projet ciblé (PC) HAMMER vient d’être officiellement accepté. C’est un PC associé au PEPR Robotique (accélération) dont les informations importantes sont données ci-après :

Résumé scientifique : 

The development of intelligent autonomous robots has long relied on model- and rule-based approaches. These consist in establishing a global model of the system, defining a set of rules to accomplish a task, and then designing a robust control law. While effective in controlled environments, such approaches quickly become inadequate as the complexity and dynamics of the real world increase, due to the sheer number of rules and parameters required. To overcome these limitations, two main strategies have been explored. The first seeks to increase the fidelity of models (geometric, dynamic, photometric), but this generally leads to an explosion in computational complexity. The second relies on data-driven approaches (machine learning), and in particular neural networks, which are capable of learning directly from large datasets. Although powerful, these methods require massive amounts of data, costly training phases, and provide no guarantee of stability or robustness in control laws. In light of these challenges, hybridization between rule-based and learning-based approaches has emerged as a promising path. The objective is to combine the analytical rigor of model-based methods with the adaptability of statistical models. This project is guided by two ambitions: (1) enriching models with knowledge extracted from data-driven methods, thereby going beyond predefined rules; and (2) constraining statistical approaches with structured knowledge derived from robotic tasks, in order to ensure physically coherent outcomes. These two axes raise major scientific challenges: the need for interpretable AI methods to capture implicit knowledge, and the design of architectures capable of embedding physical constraints. The crucial advantage of hybrid approaches lies in their ability to enhance explainability through the explicit integration of physical models. Furthermore, recent strategies exploiting Lyapunov functions—either learned or analytically derived—ensure system stability and convergence. Nevertheless, hybridization remains relatively underexplored in robotics, mainly due to the requirement for dual expertise and the absence of a consolidated theoretical framework. Our ambition is to advance the state of the art in sensor-based motion generation for complex robotic systems (drones, aerial manipulators, quadrupeds, humanoids). The research will be structured around two main axes: 

1. Development of hybrid dynamic models to improve the fidelity of simulators, system identification, and state estimation, thereby capturing complex dynamics. 

2. Integration of hybrid algorithms into control and perception strategies, with the aim of optimizing robot behavior, improving learning efficiency, and enhancing generalizability. 

These scientific and technological advances aim to establish a coherent framework for hybrid robotics, combining analytical rigor with the power of data, and thereby giving rise to a new generation of autonomous robots—robust, safe, interpretable, and capable of operating effectively in complex and dynamic environments.