Thèse IA Agentique Embarquée pour l'Assistance Contextuelle et Adaptative des Personnes Âgées à Partir de Capteurs Portés et Ambiants H/F Doctorat.Gouv.Fr

Établissement : Université Toulouse II Jean Jaurès École doctorale : GEETS - Génie Electrique Electronique,Télécommunications et Santé : du système au nanosystème Laboratoire de recherche : LAAS - Laboratoire d'Analyse et d'Architecture des Systèmes Direction de la thèse : Eric CAMPO ORCID 0000000176478498 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-23T23:59:59 Le vieillissement de la population constitue un enjeu majeur de santé publique, en particulier pour le maintien de l'autonomie, la prévention de la dépendance et l'accompagnement des personnes âgées à domicile. Dans ce contexte, les smartphones, montres connectées, capteurs portés et capteurs ambiants offrent de nouvelles possibilités pour observer de manière continue, discrète et non intrusive les activités de vie quotidienne, la mobilité, les routines et certains paramètres physiologiques.
Les systèmes actuels d'intelligence artificielle embarquée permettent déjà de reconnaître certains comportements ou événements à partir de données capteurs. Toutefois, ils restent souvent limités par des modèles statiques ou des règles prédéfinies, peu capables de s'adapter à l'évolution des habitudes, du contexte ou de l'état de l'utilisateur. Cette thèse propose de dépasser cette approche réactive en développant une architecture d'IA agentique embarquée, capable de percevoir l'environnement, d'analyser les données multimodales, de raisonner sur la situation, de planifier des actions et d'interagir de manière proactive avec la personne âgée.
L'objectif principal est de concevoir, implémenter et évaluer une application mobile intelligente destinée à l'assistance cognitive, contextuelle et adaptative de personnes âgées nécessitant une assistance modérée dans leurs activités de vie quotidienne. Le système exploitera des données issues de capteurs embarqués et portés, tels que l'accéléromètre, le gyroscope, le podomètre ou la fréquence cardiaque, ainsi que de capteurs ambiants comme les capteurs de présence, d'ouverture, d'énergie, de température ou de luminosité. Les interactions vocales seront également intégrées afin de faciliter l'usage du dispositif et de renforcer son accessibilité.
La méthodologie reposera sur la collecte de données en environnements réels et semi-contrôlés, en France et en Tunisie, puis sur le développement de modèles d'apprentissage automatique embarqués pour la reconnaissance des activités de vie quotidienne. Des mécanismes d'apprentissage adaptatif permettront de modéliser les habitudes individuelles et de détecter précocement des situations à risque, telles qu'une inactivité prolongée, une anomalie de mobilité, une fatigue inhabituelle ou un signe de fragilité. Une architecture multi-agents intégrera les fonctions de perception, d'analyse, de décision et d'interaction, dans une logique hybride combinant traitement local et analyse distante sécurisée.
Les résultats attendus concernent le développement d'une architecture complète d'IA agentique embarquée, de modèles de reconnaissance d'activités adaptés aux contraintes de calcul, d'énergie et de confidentialité, ainsi que de mécanismes proactifs d'alerte et d'assistance personnalisée. Le projet devra également produire des indicateurs synthétiques de mobilité et d'activité pouvant servir d'aide à la décision pour le suivi à distance et la prévention de la perte d'autonomie. La validation sur sites réels permettra d'évaluer la faisabilité, la pertinence et l'acceptabilité de l'approche. Population ageing is currently a major public health challenge. Increasing life expectancy is accompanied by a growing prevalence of frailty, potential declines in mobility, changes in daily routines and, in some cases, an increased risk of social isolation. In this context, maintaining autonomy and supporting ageing at home have become scientific, medical, social, and technological priorities.
Smartphones, smartwatches, and wearable devices now make it possible to continuously monitor various parameters related to physical activity, mobility, and physiological status. Embedded sensors, such as accelerometers, gyroscopes, pedometers, and heart-rate sensors, provide discreet and continuous observation of daily habits and behaviours. These devices can be complemented by ambient sensors installed within the living environment, including presence, door-opening, energy-consumption, temperature, and light sensors.
Current embedded artificial intelligence systems mainly focus on local data processing and the recognition of predefined behaviours or events. These approaches offer significant advantages in terms of privacy, processing speed, and reduced data transmission requirements. However, they often remain limited by their essentially reactive nature, their reliance on static models or fixed rules, and their difficulty in adapting to gradual changes in the user's behaviour, context, or health status.
Agentic AI represents an important evolution by introducing autonomous software entities capable of perceiving their environment, interpreting multimodal data, reasoning about a given situation, planning actions, and interacting with the user. By combining embedded AI with agentic AI, it becomes possible to develop assistance systems that are more personalised, adaptive, and proactive.
The scientific research question addressed in this PhD thesis can therefore be formulated as follows: how can an intelligent, autonomous, and embedded architecture combining perception, learning, reasoning, and interaction be designed to provide personalised, adaptive, and reliable support to older adults in their activities of daily living? The main objective of this PhD thesis is to design, implement, and evaluate an intelligent assistance system based on an embedded agentic AI architecture, intended to support older adults requiring moderate assistance in their activities of daily living.
The specific objectives are as follows:
- to design a user-centred embedded mobile application featuring a simplified, accessible interface adapted to older adults;
- to integrate a natural voice interaction module to facilitate system use and communication with the user;
- to exploit data collected from embedded, wearable, and ambient sensors in order to recognise activities of daily living;
- to develop embedded intelligent agents dedicated to perception, data analysis, and interaction with the user;
- to implement an autonomous decision-making agent combining machine learning, agentic reasoning, and context adaptation;
- to implement proactive mechanisms for detecting potentially risky situations, such as prolonged inactivity, mobility anomalies, fatigue, or early signs of frailty;
- to ensure the secure transmission of relevant data through a hybrid approach combining local embedded processing and remote analysis;
- to generate concise indicators for monitoring mobility, activity levels, and daily routines;
- to validate the system through experiments conducted in real-world and semi-controlled environments, particularly across several pilot sites. The proposed methodology is based on an experimental, technological and applied approach, combining software development, data acquisition, behavioural modelling, embedded artificial intelligence and validation under real or semi-controlled conditions.
The data used will come from several complementary sources:
- sensors integrated into smartphones;
- wearable devices, such as smartwatches or connected wristbands;
- embedded sensors: accelerometer, gyroscope, step counter, heart rate;
- ambient sensors: presence, opening, magnetic, energy, temperature and light sensors;
- data from voice interactions with the user.
A first step will consist in using open datasets, such as CASAS, and collecting experimental datasets in real and semi-controlled environments in France and Tunisia. These data will be used to characterise activities of daily living, mobility routines, variations in activity, and situations that may indicate risk or frailty.
A second step will focus on the development of embedded machine learning models for the recognition of activities of daily living. These models will need to be sufficiently accurate while complying with the constraints specific to embedded systems: limited computational power, energy consumption, robustness, low latency and privacy preservation.
A third step will address the adaptive modelling of lifestyle habits. The aim will be to identify individual routines, their evolution over time, as well as significant deviations that may indicate fatigue, frailty, unusual inactivity or a risk situation.
A fourth step will aim to design an embedded multi-agent architecture integrating several functions: perception, analysis, reasoning, decision-making and interaction. An autonomous decision-making agent will be responsible for combining the outputs of the learning models with an agentic reasoning logic, in order to propose relevant actions, reminders, recommendations or alerts.
A fifth step will focus on the development of a demonstrator mobile application. This application will integrate a simplified interface, natural voice interaction, activity monitoring functions, synthetic indicators and proactive alert or assistance mechanisms.
Finally, the system will be evaluated according to several criteria: activity recognition accuracy, relevance of alerts, ability to adapt to behavioural changes, user acceptability, energy consumption, data security and feasibility of deployment under real-life conditions.