Séminaire AID/CIEDS « Essaims de drones » et « FARO » des 10 et 11 juin 2024

Télécom Paris: le 10 juin à partir de 12:00, amphi 6, le 11 juin matin en amphi 6 également, le 11 juin après midi en amphi 3 jusqu'à 15:30.

Le social event est au Zeyer, à 7:30pm précises, le 10 juin.

Lundi 10 juin

12:00am-1:30pm déjeuner - salle 1A422 (buffet)

1:30pm-2:15pm Luc Jaulin (ENSTA Bretagne), " Validated control using interval and flatness; application to the car-trailer "

Abstract: This presentation proposes to combine methods coming from nonlinear control theory such as flatness, or feedback linearization with interval analysis, in order to guarantee that a mobile robot will not enter inside a forbidden zone. The principle is to use flatness to transform differential constraints into analytical inequalities using Lie derivatives. The resolution is performed numerically and rigorously using interval analysis tools without using any interval integration. The methodology will be illustrated on the car-trailer vehicle.

2:25pm-3:10pm, Aloysio Galvao Lopes (Ecole Polytechnique): " ConForME: Multi-horizon conditional conformal time series forecasting "

Abstract : Split conformal prediction is a statistical method known for its probabilistic coverage guarantees, simplicity, and low computational cost. As such, it is suitable for predicting uncertainty intervals in time series forecasting. However, when multiple future time steps are predicted at once, the current literature lacks conformal methods that produce tight intervals and have low computational cost. In this talk, we present ConForME, a method that builds on the foundation of split conformal prediction and one of its most prominent extensions to multi-horizon time series forecasting (CF-RNN) to construct tight multi-horizon prediction intervals with probabilistic joint coverage guarantees. In all cases, our method outperforms CF-RNN in terms of mean, minimum, and maximum interval sizes over the entire prediction horizon, achieving improvements of up to 52\%. The experiments also suggest that these improvements can be further increased by extending the prediction horizon and by hyperparameter optimization.

3:20pm-3:35pm, Break

3:35pm-4:20pm Mehdi Zadem (Ecole Polytechnique and ENSTA Paris): " Spatial and Temporal Abstractions for Goal Representation "

Abstract: Goal representation affects the performance of Hierarchical Reinforcement Learning (HRL) algorithms by decomposing the complex learning problem into easier subtasks. Recent studies show that representations that preserve temporally abstract environment dynamics are successful in solving difficult problems and provide theoretical guarantees for optimality. These methods however cannot scale to tasks where environment dynamics increase in complexity i.e. the temporally abstract transition relations depend on larger number of variables. On the other hand, other efforts have tried to use spatial abstraction to mitigate the previous issues. Their limitations include scalability to high dimensional environments and dependency on prior knowledge. We propose a novel three-layer HRL algorithm that introduces, at different levels of the hierarchy, both a spatial and a temporal goal abstraction. We provide a theoretical study of the regret bounds of the learned policies. We evaluate the approach on complex continuous control tasks, demonstrating the effectiveness of spatial and temporal abstractions learned by this approach.

4:30pm-5:15pm, Loick Degorre (ENSTA Bretagne): " Emulation du comportement d’un voilier : Commande par jumeau numérique ", and video 1, video 2, video 3, video 4

Abstract : Cette présentation décrit les récents résultats expérimentaux obtenus par l’équipe Robex de l’ENSTA Bretagne sur la problématique de l’émulation du comportement des voiliers autonomes. Afin de faciliter le développement d’algorithmes d’évitement d’obstacles pour les voiliers autonomes et l’acquisition de données capteur expérimentales, nous avons développé un système émulant le comportement d’un voilier autonome avec un vaisseau de surface autonome classique. Un voilier virtuel est simulé en temps réel et traqué par le vrai véhicule sur l’eau. Cette expérience permet notamment de faire des relevés de capteurs en condition sans n’avoir à déployer un vrai voilier autonome. L’émulation de systèmes couteux ou difficiles à déployer comme le voilier autonome permet de faciliter les développements de nouvelles méthodes de commande et d’anticiper les besoins et les résultats de ces systèmes.

Mardi 11 juin

9:30am-10:00am Eric Goubault, Laurent Pautet, Luc Jaulin, Christophe Garion, Jean-Daniel Masson. Réunion de travail

10:00am-10:45am, Bernardo Hummes Flores (Ecole Polytechnique): " Towards a sheaf theoretic view of distributed robot tasks  "

Abstract: Sheaf theory arose from the study of global properties of data described locally. This presentation will show how this point of view is compatible with the study of swarms of robots through their local interactions. The language of sheaves will be used to express distributed robot tasks and their solvability. The sheaf Laplacian construction will be presented as a generalized averaging algorithm for gathering, and a case will be made for its usage in localization and exploration tasks.

10:50am-11:35am, Nuwan Herath Mudiyanselage (ENSTA Bretagne): " New interval tools: set projection with reinforced separators  "

Abstract: In interval arithmetic, sets can be described by operators called separators. We reinforce those separators with additional information. These new operators come with an algebra. This approach is particularly useful to represent the projection of differentiable sets. Some engineering problems can be represented as problems on sets. Therefore, we focus on the projection of such sets. For instance, when the problem depends on n parameters x, y, p_3, ..., p_n, one may only be interested in the position (x,y) in R^2.

11:35am-11:50am Break

11:50pm-12:35pm Simon Rohou (ENSTA Bretagne):, " An aymptotic minimal contractor for non-linear equations, provided in the Codac library "

Abstract: We consider the set-inversion problem of a non-linear function. In particular, we will consider systems for which the solution set has infinitely many solutions. The state-of-the-art contractor for this problem is the HC4Revise algorithm. However, as for natural inclusion functions, the efficiency of an HC4Revise contractor will be impacted by multi-occurrences of the variables in the analytic expression of the function. Solutions exist to overcome this pessimism, such as symbolic rewriting or affine arithmetic, but they are not always minimal, difficult to use, or based on complicated algorithms. On the other hand, the use of centered-form computations provides asymptotically minimal results, as shown in a recent contribution from Luc Jaulin, presented during a previous FARO meeting. We propose an implementation of the centered form in the contractor framework, making use of automatic differentiation in order to automatically compute the Jacobian of the centered form. The resulting implementation will be discussed around several examples and a comparison with affine arithmetic.

12:35am-1:30pm déjeuner - salle 1A422 (buffet)

1:30pm-2:15pm (All),  Discussion on the follow up of the projects