Skip to Main content Skip to Navigation

State estimation and localization of legged robots : a tightly-coupled approach based on a-posteriori maximization

Mederic Fourmy 1 
1 LAAS-GEPETTO - Équipe Mouvement des Systèmes Anthropomorphes
LAAS - Laboratoire d'analyse et d'architecture des systèmes
Abstract : Legged robots are complex mechanisms whose stable behavior depends on the proper estimation of several different quantities that must be observed athigh speed and accuracy. While the robot state can mostly not be observed directly by any existing sensors, it is typically reconstructed by fusing verydiverse sensor modalities.These sources of information differ significantly on their acquisition frequencies, the nature of the data, and the computational processing cost.While state estimation by sensor fusion is a common topic to most robotic platforms, it offers a particular challenge in legged robotics by their peculiar dynamics.For legged robots, on the one hand, the robot state is needed to maintain its sense of balance and locomote safely, and, on the other hand, a precise representationof the environment is required for navigation and interaction.The current approach for many legged systems, in the literature, solvethese problems independently, using cascades of estimators that may neglect some of the correlation present in the data.This artificial decoupling acts as strong priors that enable simple estimators to handle the estimation of each part of the cascade and stabilizethe behavior of the overall estimation scheme. On the other hand, designing a cascade involves a lot of specialized work that hardly generalizes to newscenarios or new sensor modalities.In this thesis, we rather defend the idea of building a single tightly-coupled estimator capable of estimating all quantities needed by the robot.For this goal, the framework of a-posteriori estimation, formalized as a factor graph, is very suitable. This assertion does not come as a surprise,as factor graphs are nowadays highly popular in the SLAM literature, yet they are still under-represented in legged robots systems.In this thesis, we investigate a few avenues that we believe are crucial to achieving these goals. First, we develop tailored sensor measurement modelswith attention to the correct mathematical formulation involving Lie theory.Second, we propose visual-inertial systems based on object-level detection that provide relative transformation between the camera and the objects.We provide covariance models for two kinds of objects: the first one is an analytical model for fiducial markers; the secondis an empirical model for deep-learning-based object pose estimation.Third, we handle high-rate sensors by developing a generalization of the IMU pre-integration theory. We propose a new formulation of the IMUpre-integration based on compact Lie groups.Fourth, we show that pre-integration can also be applied to use force-torque sensors found on legged robots.By fusing it with a leg-kinematics based odometry and IMU, we show that this new formulation makes possible the tightly-coupled estimationof centroidal quantities within the context of Factor Graph estimation.The proposed theoretical ideas are implemented in a coherent estimation framework, extending the factor-graph software Wolf. Each new modality isvalidated in a dedicated experimental setup that allowed us to quantify its interest and relevance for legged robotics.
Document type :
Complete list of metadata
Contributor : ABES STAR :  Contact
Submitted on : Wednesday, July 6, 2022 - 4:54:28 PM
Last modification on : Monday, July 25, 2022 - 2:06:52 PM


Version validated by the jury (STAR)


  • HAL Id : tel-03715727, version 1


Mederic Fourmy. State estimation and localization of legged robots : a tightly-coupled approach based on a-posteriori maximization. Automatic. INSA de Toulouse, 2022. English. ⟨NNT : 2022ISAT0005⟩. ⟨tel-03715727⟩



Record views


Files downloads