Knowing a cooperative robot{\textquoteright}s relative pose is essential in multi-robot collaboration. In unknown environments, active relative pose estimation leverages front-end SLAM (Simultaneous Localisation and Mapping) or odometry with inter-robot range measurements to compute an ad-hoc observable relative pose. This dissertation uses UWB (Ultra WideBand) for inter-robot ranging and VIO (Visual Inertial Odometry) for odometry. In this setup, three to four degrees of freedom must be estimated. These are the two or three translational degrees of freedom in 2D or 3D setups, respectively, and the orientation around the gravity vector in both 2D and 3D setups. With only one UWB device per robot, the relative pose is initially unobservable but can become observable through motion. Obstructions in the Line-of-Sight (LOS) can also lead to incorrect range measurements. This work explores the feasibility of active relative pose estimation across scenarios, using an unscented particle filter to track the multi-modal pose distribution when observability is lacking. This is the first algorithm to track a 3D multi-modal solution without prior knowledge of the initial relative pose. The effect of LOS obstructions is further examined in the 2D setup. When the initial relative pose is known, a particle filter is introduced to track both the LOS state and the relative pose, improving estimation during occlusion and accelerating convergence when LOS is restored. However, the proposed solution does not generalise well over different scenarios and is very sensitive to VIO drift. To overcome this limitation, a supervised learned network is proposed to detect LOS state based on the measured odometry of both robots and UWB range, without prior pose knowledge of the initial relative pose. A simulation framework models UWB behaviour under occlusions to train classifiers, which improve LOS classification accuracy by 14\% across varied scenarios with respect to the previously developed method. By addressing underexplored challenges, this dissertation advances active relative pose estimation toward practical deployment in multi-robot collaboration. However, it was also found that the active relative pose accuracy is strongly dependent on the quality of the odometry estimation. Yet, due to its observable nature, the estimated relative pose will not drift. Conducted with the Brubotics (R\&MM) and Electronics and Information Processing (ETRO) research groups at Vrije Universiteit Brussel (VUB), this work contributes to the EU projects SPEAR and EURobin, and is supported by Flanders Artificial Intelligence Research and the imec AAA projects.
Durodié, Y 2025, 'Ad-hoc active relative pose estimation using Visual Inertial Odometry and a single UWB antenna per robot for multi-robot collaboration systems', Vrije Universiteit Brussel.
Durodié, Y. (2025). Ad-hoc active relative pose estimation using Visual Inertial Odometry and a single UWB antenna per robot for multi-robot collaboration systems. [PhD Thesis, Vrije Universiteit Brussel].
@phdthesis{b3910a9a2e3a4cbd8dcee1d1b0e949e7,
title = "Ad-hoc active relative pose estimation using Visual Inertial Odometry and a single UWB antenna per robot for multi-robot collaboration systems",
abstract = "Knowing a cooperative robot{\textquoteright}s relative pose is essential in multi-robot collaboration. In unknown environments, active relative pose estimation leverages front-end SLAM (Simultaneous Localisation and Mapping) or odometry with inter-robot range measurements to compute an ad-hoc observable relative pose. This dissertation uses UWB (Ultra WideBand) for inter-robot ranging and VIO (Visual Inertial Odometry) for odometry. In this setup, three to four degrees of freedom must be estimated. These are the two or three translational degrees of freedom in 2D or 3D setups, respectively, and the orientation around the gravity vector in both 2D and 3D setups. With only one UWB device per robot, the relative pose is initially unobservable but can become observable through motion. Obstructions in the Line-of-Sight (LOS) can also lead to incorrect range measurements. This work explores the feasibility of active relative pose estimation across scenarios, using an unscented particle filter to track the multi-modal pose distribution when observability is lacking. This is the first algorithm to track a 3D multi-modal solution without prior knowledge of the initial relative pose. The effect of LOS obstructions is further examined in the 2D setup. When the initial relative pose is known, a particle filter is introduced to track both the LOS state and the relative pose, improving estimation during occlusion and accelerating convergence when LOS is restored. However, the proposed solution does not generalise well over different scenarios and is very sensitive to VIO drift. To overcome this limitation, a supervised learned network is proposed to detect LOS state based on the measured odometry of both robots and UWB range, without prior pose knowledge of the initial relative pose. A simulation framework models UWB behaviour under occlusions to train classifiers, which improve LOS classification accuracy by 14\% across varied scenarios with respect to the previously developed method. By addressing underexplored challenges, this dissertation advances active relative pose estimation toward practical deployment in multi-robot collaboration. However, it was also found that the active relative pose accuracy is strongly dependent on the quality of the odometry estimation. Yet, due to its observable nature, the estimated relative pose will not drift. Conducted with the Brubotics (R\&MM) and Electronics and Information Processing (ETRO) research groups at Vrije Universiteit Brussel (VUB), this work contributes to the EU projects SPEAR and EURobin, and is supported by Flanders Artificial Intelligence Research and the imec AAA projects.",
author = "Yuri Durodi{\'e}",
year = "2025",
language = "English",
school = "Vrije Universiteit Brussel",
}