Description du sujet de thèse

The aim of this thesis is to propose a theoretical framework to address sobriety in robotic target tracking, by achieving the right performance required by the context via the selection of perception actions and the reduction of computational complexity.

Target tracking has been explored since the early days of data fusion. The aim is to locate targets (or objects) relative to an observer, whose number is unknown, varies and whose detection by sensors is uncertain. In robotics, tracking is a function that contributes to the spatial understanding of the environment by the decisional layer. Similarly, multi-target tracking can be found in conjunction with SLAM (Simultaneous Localization And Mapping), giving rise to the SLAM-MOT (SLAM and Moving Object Tracking) problem.

Mobile robotics adds original constraints to this problem. Calculations have to be made on-line based on present and past observations. They are subject to severe temporal requirements. In the considered variable and evolving environments, changing perceptual conditions can lead to a high rate of false observations. In addition, a mobile robot carries out its mission on batteries, so the energy available to the processing units and on-board sensors is limited. This is particularly the case for aerial robotics (quadcopters…), where the processing units, sensors, and batteries make up most of the payload (excluding specific tools). Finally, the tracking function runs concurrently with other potentially more critical functions (maintaining stability, etc.). Limiting computational complexity and energy consumption to what is strictly necessary frees up these resources for other functions. For example, ultra-precise tracking of distant targets (out of reach of the robot in the short term) is not strictly necessary.

The problem of this thesis is therefore to propose a theoretical framework to multi-target tracking that meets this sobriety requirement in robotics. The aim is to obtain a consistent and trustworthy tracking system, involving limited resources (sensors, CPU processing…), and endowed with the level of performance just necessary according to the context. It is in line with the robotics community's growing awareness of its impact in terms of resources and energy.

Conventional data fusion methods make use of all available measurements. These methods are known as “Bottom Up”. They offer the best estimation quality, but can be expensive in terms of CPU and energy resources, while some measurements provide little information. The aim of this thesis is to explore “Top Down” methods, which prioritize the most informative measurements. In this way, less strain is placed on sensors and computational resources, and the ratio between information gain and energy consumption is maximized.

In parallel, various issues will be explored: definition of metrics suited to the sober tracking problem; approximate inference and optimization strategies…

Contexte de travail

The thesis will be conducted at the Laboratoire d'Analyse et d'Architecture des Systèmes (LAAS), a CNRS unit, in the Robotics Action and Perception (RAP) team of the Robotics department (https://www.laas.fr/en/teams/rap/).

Theoretical contributions will be validated experimentally on the department's platforms.

Candidate profile: the candidate must hold a Master's or Engineering degree in one of the following fields: Robotics; Computer Science; Signal Processing; Automatic Control; Applied Mathematics.

Expected skills:
- Proficiency in scientific computing software for algorithm prototyping: Python, Julia, Matlab, etc.
- Mastery of computer tools and languages for application development and integration: Git, ROS, C++, Rust...
- Solid background in estimation/filtering and optimization.
- Fundamentals of robotics and perception.

Le poste se situe dans un secteur relevant de la protection du potentiel scientifique et technique (PPST), et nécessite donc, conformément à la réglementation, que votre arrivée soit autorisée par l'autorité compétente du MESR.