|
|
|
|
|
|
|
|
|
Pedro Alves De Araujo Silva
Dates : du 04/05/2026 au 31/07/2026
Lieu : Brest, France
Sujet : Contrôle vertical d'un drone à voilure hybride
Laboratoire : Robex, Ensta Bretagne
Encadrant : Lionel Lapierre
Résumé : This document details the mathematical modeling, kinematic transforma-
tions, and passivity-based robust control design for a 6 Degrees of Freedom (6
DOF) Stewart Platform mounted on a marine vessel. The primary objective is
to implement Active Heave Compensation (AHC) and multiaxial attitude sta-
bilization to neutralize sea wave disturbances (Roll, Pitch, and Heave). This
work addresses the critical problem of delayed feedback introduced by sensor
filtering latency and actuator response dead-time, which typically compro-
mises the stability margins of standard PID architectures. By transforming
the joint-space control loop into Wave Variables, passivity is rigorously pre-
served, guaranteeing unconditional system stability under arbitrary constant
time delays.
Rapport sur internet : Non
Jury : Léon Perruchot, Luc Jaulin, Ewen Mélé
Pedro Henrique Barbosa Nori
Dates : du 04/05/2026 au 28/08/2026
Lieu : Brest, France
Sujet : Bathymetric Data Acquisition for Micro-AUV State Estimation
Laboratoire : Robex, Lab-STICC, Ensta
Encadrant : Simon Rohou et Alain Bertholom
Résumé : This report details the work conducted during a four-month internship at the
Lab-STICC, Robex team, focusing on the implementation of a functional bathymetric
mapping pipeline for the micro-AUV Yuco (Seaber). To extend the baseline function-
ality of the AUV, a custom payload section called Yodac was introduced. It provides
a dedicated platform for testing state estimation filters, autonomous decision-making
algorithms and bathymetric mapping. The resulting system integrates an embedded
electronic board, including a single-beam echo sounder.
The core technical contribution involved designing and implementing an embedded
ROS software architecture within the Yodac payload, interfacing it with the Yuco
AUV via its communication API, and integrating the echo sounder through a custom
software library. Finally, the developments were validated through experimental sea
trials to evaluate the system’s operational behavior. Ultimately, this work serves as
a foundational starting point for future research and autonomous implementations
on the micro-AUV.
Jury : Simon Rohou, Luc Jaulin, Ewen Mélé, Erwan Le Franc
Luiz Felipe Barros Alves
Dates : du 04/05/2026 au 24/07/2026
Lieu : Brest, France
Sujet : Magnetic robot for underwater applications
Laboratoire : Ensta, Brest
Encadrant : Yannick Argouarc'H
Résumé : Improvement of the hardware architecture, optimization of the electronics, and
development of new functionalities for a magnetic robot.
This report details the hardware and software optimization of a magnetic underwater
robot. The electronic architecture was restructured (including the design of a dedicated custom
board and an upgrade of the power supply to 36 V) and new mechanical parts were integrated
using 3D printing. Concurrently, the control system was made more robust by optimizing
the interfacing of the sensors (camera, IMU), the control interface, and the brushless motors.
Finally, a prototype for detecting magnetic field variations was developed. The entire system
was validated through experimental testing in the test tank at the ENSTA Brest campus and
at the Port of Brest.
Jury : Luc Jaulin
Matheus Benigno Da Silva
Dates : du 18/05/2026 au 18/07/2026
Lieu : Brest, France
Sujet : La régulation d’un BlueBoat par apprentissage
Laboratoire ou Entreprise : Ensta Bretagne
Encadrant : Quentin Brateau
Résumé : Ce stage a permis de construire une chaîne complète de contrôle pour le BlueBoat,
depuis la modélisation en simulation jusqu’aux premiers essais réels en mer. Le travail
a d’abord consisté à définir un contrôleur par champ vectoriel capable de faire conver-
ger le bateau vers une trajectoire circulaire autour d’une bouée. Ce contrôleur a ensuite
21été utilisé comme contrôleur expert pour générer une base de données d’apprentissage
supervisé.
Une architecture de réseau de neurones compacte a été étudiée. La version principale
utilisée dans ce rapport est une architecture sans biais de type 4 − 8 − 8 − 2, composée
de quatre entrées, deux couches cachées de huit neurones avec activation ReLU, et deux
sorties correspondant aux commandes des propulseurs. La base de données finale contenait
environ 1,15 million d’exemples, obtenus à partir d’une grille de positions et d’orientations.
Le réseau a ensuite été entraîné pour reproduire les commandes générées par le contrôleur
par champ vectoriel.
Jury : Luc Jaulin
Matys Bieche
Dates : du 07/05/2026 au 11/09/2026
Lieu : Paris, France
Sujet : Développeur C++ en Navigation Robotique
Laboratoire ou Entreprise : Enchanted Tools_Paris
Encadrant : Grégoire Linard
Résumé : Dans le cadre de ma deuxième année à l’ENSTA, j’ai effectué un stage de cinq mois en
R&D Navigation Robotique chez Enchanted Tools. Ce stage a pris la forme de plusieurs
missions successives répondant aux besoins de l’équipe Navigation.
J’ai d’abord conçu un benchmark reproductible afin de mesurer les performances de na-
vigation du robot et détecter les régressions entre versions logicielles. J’ai ensuite intégré une
chaîne de relocalisation absolue par AprilTag destinée à corriger la dérive de la localisation
visuelle. Enfin, j’ai ajouté à la carte des zones polygonales typées, capables d’associer des
contraintes de navigation ou un contexte comportemental à des régions de l’environnement.
Ces travaux m’ont conduit à aborder conjointement la navigation autonome, l’intégration
dans une stack logicielle distribuée et la validation de fonctionnalités sur robot réel.
Rapport sur internet : Non
Jury : Léon Perruchot, Luc Jaulin, Ewen Mélé, Taddéo Guérin
Matéo Brun
Dates : du 04/05/2026 au 04/09/2026
Lieu : Trondheim, Norvege
Sujet : Flying a Drone with your Mind: FlexEEG motor imagery
Laboratoire : Universite Norvégienne De Science Et Technologie
Encadrant : Marta Molinas
Résumé : Ce stage de quatre mois à la NTNU m’a permis d’explorer les interfaces cerveau-
machine appliquées à la robotique, en travaillant sur la modélisation d’un drone,
l’acquisition de signaux EEG et l’entraînement de réseaux de neurones. Si les outils
nécessaires ont pu être développés, les résultats de classification restent insuffisants
pour permettre un pilotage fiable. Je regrette notamment de ne pas avoir eu le
temps de tester le système complet sur le Crazyflie réel. Des enregistrements sup-
plémentaires et une amélioration des modèles seraient nécessaires pour poursuivre
ce travail et atteindre cet objectif.
Sur le plan technique, ce projet m’a permis de découvrir un domaine pluridis-
ciplinaire et de mieux comprendre les difficultés liées à l’exploitation de données
expérimentales. Il m’a notamment appris que la réussite d’un système ne dépend
pas uniquement des algorithmes employés, mais aussi de la qualité des mesures et
de leur validation en conditions réelles.
Rapport sur internet : Oui/Non
Jury : Luc Jaulin
Noé Carrayrou
Dates : du 11/05/2026 au 11/09/2026
Lieu : Trondheim, Norvege
Sujet : Flying a drone with your mind: flex EEG motor imagery
Laboratoire ou Entreprise : Universite Norvegienne De Science Et Technologie
Encadrant : Marta Molinas
Résumé : The internship takes place at the Norvegian University of Science and
Technology (NTNU), in Trondheim, in Norway. This internship was supervised by
professor Marta Molinas, in the cybenetic department.
The research team I was in aimed at studying brain waves in a non-invasive
way, and how to use it in different fields such as medicine or robotic.
The goal of this internship was to develop a human-machine interface (HMI)
allowing us to pilot a drone thanks to measures on the brain activity, realised
by electro-encephalography (EEG). To do that, I had to develop an interface to
register the data in order to train a neural network. This network had to convert
the data into commands for the drone. When the network is sufficiently trained, I
had to develop a flight simulator to train to pilot the drone before flying the drone
for real.
Jury : Luc Jaulin
Elodie Conan
Dates : du 04/05/2026 au 04/09/2026
Lieu : Kongens Lyngby, Danemark
Sujet : Enhancing Situational Awareness for Maritime Autonomous Surface Ships (MASS) with RGB and LWIR camera fusion
Laboratoire : Technical University Of Denmark (Dtu)
Encadrant : Morgan Louédec, Roberto Galeazzi, Jonathan Eichild Schmidt
Résumé : Maritime autonomous surface ships (MASS) require reliable situational awareness, particularly
in low-light and adverse weather conditions where visible-spectrum cameras may be limited. This
internship, conducted within the CREA research group at DTU in Copenhagen, investigates the use
of a long-wave infrared (LWIR) camera combined with a pair of RGB cameras for the detection,
localisation, and tracking of boats from a fixed land-based platform.
Boats are detected and segmented in LWIR images using a YOLO11-seg model fine-tuned on the
MassMIND dataset. The position of each boat is then estimated by intersecting the corresponding
viewing ray associated with the detected waterline pixel with the sea plane. These estimates are fused
with position estimates from an RGB stereo branch using a world-frame Kalman filter. The approach
was evaluated on a sequence recorded in Copenhagen Harbour. The results demonstrate the feasibility
of localising and tracking boats by combining thermal and RGB vision. The developed framework
provides a foundation for future research at DTU towards further improving boat localisation through
the integration of thermal cameras.
Jury : Luc Jaulin
Gaspard Delpiano Manfrini
Dates : du 04/05/2026 au 11/09/2026
Lieu : Tsukuba, Japon
Sujet : Purpose of Technical Training: To learn technologies related to deep-learning-based robot motion planning and computer vision through reasearch work.
Laboratoire ou Entreprise : National Institute Of Advanced Industrial Science And Technology
Encadrant : Qiu Yue
Résumé : This internship report presents the work done on the implementation of sound in a
Vision Language Action based control policy, an end-to-end framework for sound and
vision based robotics manipulation, in the Computer Vision Research Team at the
Artificial Intelligence Research Center (AIRC), part of Japan’s National Institute of
Advanced Industrial Science and Technology (AIST) in Tsukuba, Japan. A sound VLA
model was developed based on the architecture and weights of SmolVLA. The base
architecture was expended upon by adding a Mel Spectrogramme encoder and training
a sound projector.
The work also includes a physics and sound simulation environment leveraged for cus-
tom dataset generation as well as in the reinforcement learning and inference testing
pipeline.
Although finalising the training and testing of the model on Benchmarks has not been
completed due to a lack of time, this work is still ongoing.
Jury : Luc Jaulin
Kadijatou Diallo
Dates : du 04/05/2026 au 04/09/2026
Lieu : Lisbon, Portugal
Sujet : Développement d’un système de surveillance sous-marine par vision et intelligence artificielle.
Laboratoire ou Entreprise : Blue Oasis
Encadrant : Giacomo Dieci
Résumé : This internship at BlueOasis allowed me to work on three complementary components of an underwater
monitoring system: the mechanical design of a drop camera, a stereo vision system for metric measurements,
and an image enhancement step for fish detection.
The stereo vision work showed me that a reliable 3D measurement depends more on the quality of the
calibration and its validation than on the visual appearance of the reconstruction. I also learned to build a
complete pipeline and to understand how errors propagate from one stage to the next.
The drop camera taught me to design a structure with incomplete information by stating clear assump-
tions, while taking into account corrosion, hydrodynamics and biofouling. The detection experiments showed
that a visually better image does not necessarily improve detection: applying WaterNet only during the
fine-tuning of RF-DETR lowered the mAP@50:95 from 0.5964 to 0.5443, so each step must be tested on the
target data.
Jury : Luc Jaulin
Luíza Diesel Chesani
Dates : du 04/05/2026 au 03/08/2026
Lieu : Brest, France
Sujet : Étude et cartographie du champ magnétique au bassin d’expérimentation air-mer à l’aide de robots de surface DDBoat
Laboratoire ou Entreprise : Airbus Operators Gmbh
Encadrant : Natacha Caouren , Luc Jaulin
Résumé : Conducted at ENSTA within STICC under the supervision of Luc Jaulin, this internship
contributed to preparing DDBoats for teaching in 2026–2027. It investigated magnetic field
variations at the Air-Sea Experimental Basin using DDBoat surface robots. The work included
platform preparation, magnetometer calibration, development of control and logging software,
remotely controlled traverses in two directions, and processing of GPS and magnetic measure-
ments. The resulting plots show spatial variations and differences between passes. However,
GPS accuracy and trajectory adjustment limit the localization of anomalies. The resulting map
is therefore an exploratory product that requires more accurate positioning and repeated mea-
surements before it can support reliable navigation.
Jury : Natacha Caouren , Luc Jaulin
Hugo Doré
Dates : du 04/03/2026 au 07/08/2026
Lieu : St Germain En Laye, France
Sujet : Étude et conception d’un réseau de neurones pour la navigation magnéto-inertielle
Laboratoire ou Entreprise : Exail Sas 78
Encadrant : Athiththan Kumarathasan
Résumé : Ce rapport présente le développement complet d'un système d'estimation de cap magnétique fondé sur un
réseau de neurones, depuis la conception de l'algorithme jusqu'à son portage sur cible embarquée.
L'objectif opérationnel est de fournir un cap fiable dès le démarrage d'un porteur, à l'arrêt et sans signal
satellite, afin d'initialiser des centrales inertielles non gyrocompassantes ou d'accélérer l'initialisation des
centrales gyrocompassantes.
Le coeur du système est un perceptron multicouche (MLP) qui apprend la déformation magnétique locale
propre à l'environnement d'installation — l'effet combiné des distorsions dites hard iron et soft iron, ainsi
que les perturbations d'ordre supérieur. Le réseau prend en entrée les trois composantes du champ
magnétique et l'attitude du porteur, et restitue un cap sous forme circulaire assorti d'une estimation
d'incertitude.
Jonathan Escoffier
Dates : du 04/05/2026 au 28/08/2026
Lieu : La Spezia, Italie
Sujet : data analysis and signal processing applied to multi-sensor detection of floating objects
Laboratoire ou Entreprise : Cmre - Centre For Maritime Research And Experimentation
Encadrant : Metin Pekkaptan
Résumé :
Daniel Farlei Silva Souza
Dates : du 04/05/2026 au 31/07/2026
Lieu : Brest, France
Sujet : Contrôle vertical d'un drone à voilure hybride
Laboratoire ou Entreprise : Ensta Bretagne
Encadrant : Lionel Lapierre
Résumé : This document details the mathematical modeling, sizing, and dynamic
validation of an Unmanned Aerial Vehicle (UAV) in an autorotation state
(autogyro descent) and active flight. The approach establishes the me-
chanical design specifications, proceeds through the aerodynamic sizing of
the main rotor and control surfaces, and derives the 3 Degrees of Freedom
(3 DOF) Ordinary Differential Equations (ODEs). The phenomenological
model is rigorously parameterized using Blade Element Theory (BET) data
extracted from a Simscape physical simulation, and the numerical ODEs are
ultimately validated against the multi-body physics engine.
Rapport sur internet : Non
Jury : Lionel Lapierre, Luc Jaulin
Vianney Fleury
Dates : du 01/01/2026 au 31/06/2026
Lieu : Bassin d’Arcachon
Sujet : Développement de solutions de navigation autonome pour bateaux de plaisance hybrides
Entreprise : BlueNav
Encadrant : Matthieu Bagory
Résumé : Ce rapport présente les travaux menés au cours d’un stage d’ingénieur réalisé entre janvier et juin 2026
au sein de l’entreprise BlueNav, spécialisée dans l’hybridation de bateaux de plaisance et basée sur le bassin
d’Arcachon. L’objectif principal du stage était le développement de la fonctionnalité d’Autodocking : un
système permettant la reproduction automatique d’une manœuvre d’accostage préalablement enregistrée
par l’utilisateur, en s’appuyant exclusivement sur des données GPS.
Après une présentation de BlueNav et de son environnement technique, le rapport détaille trois contributions préliminaires
ayant préparé le terrain à l’Autodocking : l’exploration d’une architecture de PIDs
cascadés pour le contrôle de vitesse, la correction du Control Allocation pour compenser un défaut à la
montée en régime des moteurs, et la fiabilisation du mode No Drift par projection du COG depuis l’antenne
GPS vers le COLR du bateau.
La partie centrale décrit l’architecture de l’Autodocking, articulée autour de deux modes complémentaires : le mode Transit, basé sur un algorithme Line Of Sight adaptatif gérant la phase d’approche à
vitesse modérée, et le mode Docking, basé sur un asservissement PID triple en position pour la phase
finale à basse vitesse. La gestion de la transition entre ces deux modes, avec hystérésis et contrainte sur le
point de visée, est également présentée
Rapport sur internet : Non
Jury : Léon Perruchot, Luc Jaulin, Ewen Mélé, Taddéo Guérin
Letícia Furuta
Dates : du 07/05/2026 au 31/07/2026
Lieu : Valladolid, Espagne
Sujet : Migrating a Laparoscopic Surgical Robot System from ROS1 to ROS2
Laboratoire : University Of Valladolid - Institute Of Advanced Technologies For Production - Espagne
Encadrant : Ana Cisnal De La Rica
Résumé : This report details the work conducted during a three-month internship at the
University of Valladolid, in the ITAP medical robotics laboratory. The work focused on
migrating a laparoscopic surgical robot system, originally developed by the research group
in ROS1, to ROS2. Following the end of official support for ROS1, the system (although
fully functional) had begun to grow obsolete, which motivated its port to a maintained,
long-term-support distribution.
The core of the internship consisted of studying the original codebase to understand
the system, and then rewriting and restructuring it for ROS2. The system is based on a
UR3e robotic arm equipped with a force/torque sensor and camera-based tool detection.
The migration was also an opport
Rapport sur internet : Oui/Non
Jury : Lionel Lapierre, Luc Jaulin
Louis Habigand
Dates : du 06/05/2026 au 04/09/2026
Lieu : Gent, Belgique
Sujet : Contribution to Sensor Validation and Oceanographic Data Collection Campaigns
Laboratoire ou Entreprise : Ghent University
Encadrant : Griet Neukermans
Résumé : This internship was carried out within the Marine Optics and Remote Sensing
(MARSENS) laboratory at Ghent University. The group develops innovative optical instru-
ments and observation methods to improve the understanding of marine biogeochemical pro-
cesses and the ocean carbon cycle. The main objective of the internship was to contribute to
the preparation and execution of an oceanographic campaign supporting the field validation
of the LISST-PIC sensor, an instrument specifically designed to measure particulate inorganic
carbon concentrations in seawater. The work performed involved engineering activities, field
operations, data processing and scientific literature review. A major part of the internship
focused on the design and preparation of an instrument cage intended to support multiple
optical sensors during deployments down to 2000 m depth aboard the research vessel Belgica.
In addition to research activities, the internship provided the opportunity to participate in
the Ocean Optics XXVII conference, one of the leading international conferences dedicated
to marine optics and ocean-colour remote sensing. Contributing to the organisation of the
event and interacting with researchers from around the world provided valuable exposure to
the international scientific community. Beyond technical contributions, this internship provided
valuable experience in marine optics, oceanographic field work, scientific instrumentation and
multidisciplinary research related to the marine carbon cycle.
Jury : Luc Jaulin
Simon Klécha
Dates : du 11/05/2026 au 31/07/2026
Lieu : Bucharest, Roumanie
Sujet : Image Sensor Based Odometry
Laboratoire : Academia Technica Militara Ferdinand 1 - Bucharest - Roumanie
Encadrant : Cristian Molder
Résumé : This report outlines the work I carried out during my placement as an engineering assistant
at the MTA (Ferdinand I Military Technical Academy) in Bucharest, Romania, as well as the
challenges involved. Indeed, visual odometry is currently experiencing a boom in robotics, due to
its low cost combined with acceptable efficiency and the unstoppable expansion of robotics itself.
Yet this technology has been part of everyone’s daily life for decades, particularly through its use in
computer mice. As an engineering student at ENSTA (Ecole Nationale Superieure de Techniques
Avancées) specialising in autonomous robotics, I worked on implementing this type of odometry
using the example of a small terrestrial robot called Nexus2, developed by the Academy’s robotics
laboratory. Equipped with two optical sensors from computer mice, it was initially the subject of a
theoretical study on my part aimed at establishing a control scheme for trajectory tracking, which
took both sensors into account in order to compensate for their inability to measure rotation. In
particular, I adapted Dubin’s car model, which perfectly models the locomotion of the Nexus2.
However, the challenge of actual implementation faced numerous obstacles. The robot is old and
there have been persistent software compatibility issues. Furthermore, my limited knowledge of
microcontrollers caused me to waste time, and I was unable to access all of the robot’s components.
Although I was unable to assess the technology’s effectiveness, this did not prevent this placement
from confirming how straightforward visual odometry is to use, at least in terms of theoretical
control.
Anatole Le Boudec Molenaar
Dates : du 07/05/2026 au 10/09/2026
Lieu : Trondheim, Norvege
Sujet : Automatic dream emotion recognition during REM sleep based on minimally invasive EEG
Entreprise : Universite Norvegienne De Science Et Technologie - Trondheim-Norvege
Encadrant : Marta Molinas
Résumé : Dream emotions remain poorly understood, yet they provide a unique opportunity to study
emotional processing during sleep, when the brain is largely disconnected from external
stimuli. This report presents two complementary contributions toward the automatic recog-
nition of dream emotions from minimally invasive EEG recordings. First, a sleep-stage
classification system was developed and evaluated on the Sleep-EDF Expanded database,
containing recordings from 153 subjects. The proposed approach achieved performance
comparable to recent single-channel EEG methods. Second, dream emotion recognition was
investigated using the REM_Turku dataset, which includes REM sleep recordings from 18
subjects. A rigorous subject-independent validation strategy was used to assess whether the
identified EEG patterns could generalize to new individuals. The results show that a reliable
EEG signature associated with anger can be identified, whereas the relationships observed
for other emotions do not generalize consistently across subjects. Overall, these findings
highlight both the potential of EEG for studying dream emotions and the methodological
challenges involved in developing reliable subject-independent emotion recognition systems.
Jury : Luc Jaulin
Juliette Lourdault
Dates : du 04/05/2026 au 06/09/2026
Lieu : Bangkok, Thailande
Sujet : Conception d’une architecture neuronale hybride à fusion multimodale pour la classification automatique d’événements sismiques volcaniques
Laboratoire : Kasetsart University - Bangok - Thailande
Encadrant : Piya Limcharoen
Résumé : Automated seismic monitoring of active volcanoes increasingly relies on deep learning
methods able to classify, continuously and without human intervention, the different types
iiiof events recorded by monitoring networks. This internship, carried out at the Department
of Computer Engineering of the Faculty of Engineering of Kasetsart University, focused
on the automatic classification of four types of seismic events from the Llaima volcano
(Chile) : long-period (LP), tremor (TR), volcano-tectonic (VT) and tectonic (TC), from
raw signals sampled at 100 Hz over 60-second windows.
The work first consisted in re-implementing three reference methods from the literature :
a convolutional classifier applied to spectrograms (Curilem et al., 2018), a comparison of
deep convolutional and recurrent architectures (Canário et al., 2020), and a multilayer-
perceptron classifier applied to temporal and spectral envelope descriptors optimised with
a genetic algorithm (Ait Laasri et al., 2025). These works served as the foundation for an
iterative design process, documented in this report in five stages : from a simple-average
CNN + BiLSTM ensemble, to a grid-search-optimised weighted ensemble, to end-to-end
hybrid architectures (CNN→BiLSTM and BiLSTM→CNN), the integration of a second
raw-signal input, and finally the addition of a non-linear multilayer-perceptron fusion
block between the two branches. This final architecture, trained and evaluated through
5-fold stratified cross-validation on the 3,592 signals of the dataset, constitutes the original
contribution of the internship and was the subject of a scientific article. With only 808,036
parameters, it reaches an accuracy of 97.77 % ± 0.41, a macro F1-score of 96.77 % ± 0.41
and a Cohen’s kappa of 0.967, outperforming the three re-implemented baseline methods
(CWT + CNN: 95.85 %; CWT + BiLSTM: 97.44 %; envelopes + MLP: 95.60 %). The
margin is larger in macro F1 than in accuracy, indicating that the gain from bimodal
fusion is concentrated on the minority classes (VT, TR).
Jury : Luc Jaulin
Thibault Lucas
Dates : du 04/05/2026 au 11/09/2026
Lieu : Tokyo, Japon
Sujet : Stage de recherche en vision par ordinateur et deep learning appliqués à la robotique
Laboratoire ou Entreprise : National Intitute Of Advanced Industrial Science And Technology-Artificial Intelligence Research Center - Tokyo
Encadrant : Qiu Yue
Résumé : Visual and pose data alone are sometimes limited for robotic control models to accomplish some manipulation tasks. Hence, the objective of our research
was to investigate the addition of a new modality, sound, in neural network based control methods.
This report presents the development of a framework for a vision language action (VLA) model in which a sound comprehension component
was added alongside the visual, pose and textual inputs. In addition, it presents the development of a data collection, training and testing pipeline.
This pipeline can be applied in simulation using a modified version of RLBench, or in reality using a Franka Panda robotic arm experimental setup. The resulting pipeline was used to
train and deploy a first sound-augmented policy based on SmolVLA in our simulation, thereby demonstrating the experimental functionality of our framework. The project resources can be
found on the following GitHub repository :
github.com/LeChatduCheshire/AIST_audio_VLA.
This work was done in the context of a four month research internship at AIST, Tsukuba, Japan. The report will also detail the learning and adaptation process I went through
to undertake this project in what was for me a new cultural and technical context.
Maël Mabilotte
Dates : du 04/05/2026 au 04/09/2026
Lieu : Patras, Grece
Sujet : Algorithm swarm localization sensors state estimation mining robots
Laboratoire ou Entreprise : University Of Patras
Encadrant : Panagiotis Koustoumpardis
Guilhem Merlet
Dates : du 04/05/2026 au 04/09/2026
Lieu : Trodheim, Norvege
Sujet : Flying a Drone with Your Mind: FlexEEG Motor Imagery
Laboratoire : Norwegian University Of Science And Technology (Ntnu) - Trondeim
Encadrant : Marta Molinas
Résumé : This internship report presents the work carried out within the Department of Cybernetics at
the Norwegian University of Science and Technology (NTNU), as part of the second year of the
engineering programme at ENSTA. The placement builds on the master’s thesis work of Daniel
Nicolai Grønn, who developed an experimental platform for controlling Crazyflie 2.1 drones using
a brain-computer interface (BCI) based on EEG motor imagery.
The main objective was to revisit this platform, rectify its malfunctions and improve its relia-
bility, so that it could serve as a solid foundation for evaluating BCI control strategies. This work
involved restoring the Loco Positioning System (LPS), resolving a critical compatibility issue be-
tween the anchor communication modes, and then fine-tuning the PID controller parameters to
stabilise the drones’ flight, both in altitude and in the horizontal plane. A rigorous experimental
approach, based on the analysis of real flight data, enabled the identification of settings offering
the best compromise between convergence speed, stability and positioning accuracy.
Finally, comprehensive documentation (User Guide Version 3.0) and a practical demonstra-
tion were produced to facilitate the transfer of this knowledge to future interns, who will continue
this project by incorporating more advanced EEG classification methods.
Rapport sur internet : Oui
Jury : Léon Perruchot, Luc Jaulin, Ewen Mélé
Evann Michel
Dates : du 04/05/2026 au 28/08/2026
Lieu : Rosieres Pres Troyes, France
Sujet : La robotique pour l’optimisation en entrepôt
Entreprise : Sequence Robotic - Rosieres Pres Troyes - 10
Encadrant : Gilles Delaunay
Résumé : This report details a four-month work internship at Sequence Robotics, a company
based in Troyes (France). The internship focused on the implementation of robotics in
warehouse environments in various ways: the use of a robotic arm for the assembly-line
production of parts, the use of autonomous AGV robots for the transport and sorting
of objects, and the development and design of modules to simplify and optimise these
solutions.
Florian Milin
Dates : du 04/05/2026 au 04/09/2026
Lieu : Bangkok, Thailande
Sujet : Development of machine learning models based on sensor data with a particular emphasis on deep learning approaches.
Laboratoire ou Entreprise : Kasetsart University - Bangok - Thailande
Encadrant : Piya Limcharoen
Résumé : This report describes my second-year internship at Kasetsart University (Bang-
kok) on vehicle counting from images with the TRANCOS dataset. After re-
viewing density-map regression methods, we re-implemented CCNN, Hydra-3s
and CSRNet, whose results differ from the published ones. We then tested seve-
ral improvements : adaptive-kernel density maps and vehicle-centred patches for
CCNN, and a multi-scale back-end and a combined loss for CSRNet. Finally, Ti-
nyNet, a lightweight architecture of 0.19 million parameters that predicts a point
map, reaches a GAME(0) of 3.45 on the test set, close to the value published
for CSRNet (3.56) with 85 times fewer parameters. This work led to a scientific
paper written in English.
Rapport sur internet : Oui
Jury : Léon Perruchot, Luc Jaulin, Ewen Mélé
Alexandre Pinard
Dates : du 07/05/2026 au 10/09/2026
Lieu : Trondheim, Norvege
Sujet : Automatic sleep stage classification based on minimally invasive EEG
Laboratoire ou Entreprise : Universite Norvegienne De Science Et Technologie
Encadrant : Marta Molinas
Résumé : This internship, carried out at the Department of Engineering Cybernetics of the Norwegian
University of Science and Technology (NTNU), focuses on automatic sleep stage classification
from minimally invasive EEG, in collaboration with the IIIS Human Sleep Lab at the University
of Tsukuba, and builds on the host laboratory’s Dream Decoder research project, which aims
to automatically decode the mental content of sleep. The work is organized into three levels of
increasing difficulty: sleep stage classification, dream presence detection, and dream emotional
tone characterization.
At the first level, a CNN + LSTM pipeline trained on single-channel EEG (Sleep-EDF
Expanded dataset, 153 subjects) reaches a Cohen’s Kappa of 0.83, an excellent level of agreement,
validating the feasibility of high-quality automatic sleep scoring on a reduced montage. At the
second level, dream detection is studied across three datasets (Multiple Awakenings, Tononi,
Zhang & Wamsley) using a PSD + XGBoost pipeline with permutation-based channel selection ;
a Kappa around 0.80 is obtained under standard cross-validation but collapses to near zero under
Leave-One-Subject-Out (LOSO) validation, revealing a systematic subject-level leakage that this
internship documents and quantifies across every dataset tested.
At the third and most difficult level, classifying anger reported in dreams (REM_Turku
dataset), based on three frontal gamma-power features selected through correlation analysis
with multiple-comparison correction (FDR), reaches 61% LOSO accuracy and proves statistically
significant under a permutation test (p = 0.048) – the only one of twenty tested emotions to
pass this check. This targeted success contrasts with the absence of significant signal for a
broader positive / negative valence classification, and with the failure to replicate the historical
correlation between frontal alpha asymmetry and anger (Sikka et al., 2019): EEG signatures of
dream emotion appear to exist, but specifically rather than generally.
Beyond the numerical results, this internship documents and corrects several classic method-
ological pitfalls of EEG-based machine learning on small samples: subject-level leakage between
training and test sets, non-physiological artefacts confused with the signal of interest, labelling
errors, and optimistically tuned decision thresholds. This methodological rigor, applied con-
sistently across all three parts of the report, is what makes it possible to separate genuinely
defensible results from those that are not.
Jury : Luc Jaulin
Lancelot Ramis-Cladera
Dates : du 11/05/2026 au 04/09/2026
Lieu : Ploudalmezeau, France
Sujet : Développement d’un jumeau numérique pour machines HVAC
Laboratoire ou Entreprise : Energie Transfert Thermique
Encadrant : David Taloc
Résumé : This internship focused on the development of a digital twin for heating, ventilation and air-
conditioning equipment at Énergie Transfert Thermique. The objective was to let a real pro-
grammable controller drive a simulated HVAC unit and building, so that control strategies could
be observed before a complete physical machine was made available. The existing repository
already contained an Electron application, a React user interface, HTTP and Modbus connectors,
an input/output engine and three elementary simulations. These simulations were directly bound
to I/O indexes and wall-clock time; the work extended this foundation into a coherent air-handling
and refrigeration chain.
Jury : Luc Jaulin
Clément Sartre
Dates : du 04/05/2026 au 05/09/2026
Lieu : Bangkok, Thailande
Sujet : AI X Finance Models to predict stock return
Laboratoire : Kasetsart University - Bangok - Thailande
Encadrant : Miti Ruchanurucks
Résumé : This report presents my second-year internship at the Computer Engineering department of
Kasetsart University, in Bangkok, whose goal was to study to what extent machine learning can
predict the future returns of stocks.
I first reproduced two articles that predict daily returns from price history alone (Abol-
makarem et al., 2024; Prasetyo et al., 2025): in both cases, my out-of-sample R2 stays close to
zero or negative.
I then attempted to reproduce the article of Gu, Kelly and Xiu (2020) with public data
only (Yahoo Finance, SEC XBRL filings and the Kenneth French Data Library), on about 500
S&P500 stocks, 21 characteristics and nine yearly test windows from 2017 to 2026. The best
model, OLS3, reaches an out-of-sample R2 of about 1.6 %, but a constant forecast alone already
reaches 1.50 %: its three characteristics only add about 0.1 percentage point, mostly through the
size effect.
These results show that the signal contained in stock data is thin compared with the noise of
returns, and that a strict out-of-sample protocol and leakage checks are essential before trusting
a predictive performance.
Rapport sur internet : Oui
Jury : Léon Perruchot, Luc Jaulin, Ewen Mélé
Oscar Verilhac
Dates : du 01/05/2026 au 31/07/2027
Lieu : Tsukuba, Japon
Sujet : Adaptation dynamique et contrôle en couple du corps entier humanoïde
Laboratoire ou Entreprise : Cnrs - Aist- Joint Robotics Laboratory - Irl 3218 _ Tsukuba _ Japon
Encadrant : Mehdi Benallegue
Résumé :
Tom Voisine
Dates : du 04/05/2026 au 04/09/2026
Lieu : Grimstad, Norvège
Sujet : Robotics electronic waste disassembly automation ROS2 software hardware interface system integration testing collaboration.
Laboratoire ou Entreprise : Entreprise Norce Norwegian Research Centre
Encadrant : Atle Aalerud
Résumé : This internship, carried out at the CARM Center of the NORCE research institute
in Grimstad, Norway, focuses on the development of an autonomous robotic system for
opening end-of-life computers (laptops), in the context of recovering critical raw materials
and valuable components from electronic waste. The objective is to develop a solution
capable of opening the devices before shredding, while preserving their components and
reducing the need for manual operations.
The system is based on a UR10e robotic arm controlled using ROS 2 and MoveIt
2, together with two interchangeable tools: a spindle equipped with a drill bit to drill
the screw heads and a vacuum gripper to remove the cover located on the back of the
device. A fixed camera provides the positions of the screws and the gripping point.
A software architecture based on several ROS 2 services was developed to control the
robot movements, spindle, gripper, and tool changes. Specific contact detection and force
control functions were also developed to improve the accuracy and safety of the drilling
process.
The proposed approach aims to go beyond conventional sequential unscrewing by di-
rectly drilling the screw heads. This method offers several advantages, including better
tolerance to damaged screws and lower positioning accuracy requirements for the tool.
The tests carried out show that the laptop can be mechanically opened using this ap-
proach. However, the sequence remains mainly sequential. Further research on the sub-
ject is still needed to consider direct industrial use in the future. Nevertheless, this work
provides a reusable hardware and software architecture for future developments aimed at
improving the speed and autonomy of the disassembly process.
Jury : Luc Jaulin
Eliaz Yahia
Dates : du 04/05/2026 au 28/08/2026
Lieu : La Spezia, Italie
Sujet : Renforcement learning applied to path planning
Laboratoire : Cmre - Centre For Maritime Research And Experimentation
Encadrant : Vincenzo Laveglia
Jury : Luc Jaulin