Estimation d'état et odométrie basées modèle pour robots à pattes

Legged robots, such as bipeds and quadrupeds, are under-actuated systems: they lack actuators dedicated to directly modifying their global position or orientation in space. Their motion instead is generated through reaction forces at contact points with the environment, which are themselves produced through the modulation of body shape and internal dynamics. Furthermore, the dynamics and kinematics of legged robots constrain the admissible postures and trajectories that allow them to maintain balance. It is therefore essential to know these states as precisely as possible to ensure the robot's stability and mobility in complex environments. This capability is indispensable for their deployment in real-world, human-centered applications.However, these quantities cannot always be measured directly, or at least not with sufficient accuracy, latency, or reliability under all conditions. It is therefore necessary to estimate the robot's state by exploiting, as coherently as possible, both the available sensor measurements and the robot's physical model. The objective is to achieve the best balance between computational efficiency, estimation accuracy, and robustness to perturbations as well as to sensor and modeling uncertainties.Building upon this analysis, this thesis introduces new state estimators that advance legged robot mobility along two complementary directions. The first focuses on high-frequency, low-latency feedback, providing strong stability guarantees and accurate estimation across a wide range of motions. The second leverages a richer physical model of the robot, further enhancing robustness to modeling errors and enabling more reliable and adaptive physical interaction with the environment. Together, these approaches establish a unified, theoretically grounded, and experimentally validated framework for state estimation in legged robot mobility, capable of supporting robust and adaptive control across diverse and challenging real-world scenarios.