Heudiasyc, UMR CNRS7253-université de technologie de Compiègne (UTC), France. Research team : SyRi website : https: // www. hds.utc.fr/
Context The context of this thesis is the navigation of autonomous vehicles. These are vehicless that can localize themselves in their environment, perceive it, interpret it without any human interaction, interact with other robots by exchanging relevant information and that make safe decisions to achieve their mission. However, so far there is no solution able to operate autonomously in different environments without human intervention. The integrity problem leading to the safety of mobile vehicles hinders their deployment. The notion of integrity is very important when autonomous navigation presents high safety risks where the probability of accident should be reduced as much as possible as in the case of autonomous vehicles. To do so, navigation requires an accurate and consistent estimation of the pose. However, the notion of integrity is largely correlated to the estimation method and the adopted assumptions. This PhD thesis is part of the ANR-JCJC project ToICaR. A postdoctoral will work on this project and a close collaboration will be carried out with this PhD to extend the approach to a collaborative multi-robot system.
Summary of the subject This thesis aims to investigate the safety of autonomous robots by dealing with the integrity of the estimates. However, this notion is largely correlated to the estimation method and the adopted assumptions. Indeed, for a system with multiple sensors, the state estimation problem is realized through a multi-sensor data fusion often done with a Kalman filter (KF). This filter, which is well suited for real-time implementation, applies to linear or linearized systems under the assumption of white Gaussian noise. The methods developed in this thesis will try to improve estimation by avoiding classical assumptions while keeping a real-time aspect. Bayesian approaches will be studied at first, as they are well adapted to statistical studies. We have recently studied the use of the Student's t-filter for multi- sensor fusion in a noisy environment 1. This study will be pursued in this thesis while dealing with the non-linearity problem. On the other hand, the coupling of several estimation methods makes possible to take advantage of each of them. Coupling Bayesian approaches and set membership representations is a good strategy for dealing with random and deterministic systematic errors 2 3. For autonomous robots, our goal is to profit from both representations in order to deal with the problem of underestimating the uncertainty value, which usually occurs with KF and overestimation, which usually occurs with the set membership representation. On the other hand, the contribution of the parameter learning strategy (covariance matrix, degrees of freedom, fault detection threshold, fault amplitude) to the state estimation method will be investigated in terms of accuracy and integrity. Moreover, one of our goals is that the learning approach can deal with partial ground truth. Therefore, a latent variable model with EM inference 4 or an elliptical process latent variable model can be used 5. However, we believe that maintaining a parametric model of the system, with known motion and observation models, will help understanding its physical behavior and it will simplify the learning step. This will lead to new adaptation of existing methods.
Even if the estimation method plays a crucial role in the design of a high integrity system, these methods are not able to manage some type of errors in an autonomous way (resulting from erroneous sensor measurements). For this reason, a fault tolerance aspect should be added. While we have some expertise in the field of fault detection and isolation 6, a special focus will be given to fault identification. Therefore, we intend to estimate the magnitude of the fault using methods that rely on EM. The purpose is to correct the faulty measurement to continue to use it when possible. Experimental part
The thesis will be based on real data to be acquired on the experimental vehicles of the Heudiasyc laboratory. These vehicles are equipped with different sensors: GNSS (Global Navigation Satellite System) receivers, wheel speed sensors, Lidars, standard and event-based cameras. High-definition maps are also available.
1 J. Al Hage, P. Xu, and P. Bonnifait, “Student's $ t $ Information Filter with Adaptive Degree of Freedom for Multi-Sensor Fusion,” in 2019 22th International Conference on Information Fusion (FUSION), 2019, pp. 1–8. 2 L. Sun, H. Alkhatib, B. Kargoll, V. Kreinovich, and I. Neumann, “A new Kalman filter model for nonlinear systems based on ellipsoidal bounding,” arXiv preprint arXiv:1802.02970, 2018. 3 C. Combastel, “Merging Kalman filtering and zonotopic state bounding for robust fault detection under noisy environment,” IFAC-PapersOnLine, vol. 48, no. 21, pp. 289–295, 2015. 4 N. Sammaknejad, Y. Zhao, and B. Huang, “A review of the Expectation Maximization algorithm in data-driven process identification,” Journal of Process Control, vol. 73, pp. 123–136, Jan. 2019, doi: 10.1016/j.jprocont.2018.12.010. 5 J. Ko and D. Fox, “Learning GP-BayesFilters via Gaussian process latent variable models,” Autonomous Robots, vol. 30, no. 1, pp. 3–23, 2011. 6 J. Al Hage, M. E. El Najjar, and D. Pomorski, “Multi-sensor fusion approach with fault detection and exclusion based on the Kullback–Leibler Divergence: Application on collaborative multi-robot system,” Information Fusion, vol. 37, pp. 61–76, Sep. 2017, doi: 10.1016/j.inffus.2017.01.005.Additional comments
Possibility of teaching at UTC Some stays at Imperial CollegeWeb site for additional job details
https: // emploi.cnrs.fr/Offres/Doctorant/UMR7253-JOEALH-003/Default.aspxRequired Research Experiences
Engineering: Master Degree or equivalent
Computer science: Master Degree or equivalent
Mathematics: Master Degree or equivalent
FRENCH: BasicContact Information