Thèse Interaction Physique Personnalisée et Proactive Entre l'Humain et le Robot Grâce à la Prédiction de Mouvements Humains H/F - Doctorat.Gouv.Fr

Établissement : Université de Toulouse École doctorale : SYSTEMES Laboratoire de recherche : LAAS - Laboratoire d'Analyse et d'Architecture des Systèmes Direction de la thèse : Bruno WATIER ORCID 0000000331462884 Début de la thèse : 2027-09-01 Date limite de candidature : 2026-11-23T23:59:59 Les robots sont de plus en plus envisagés pour assister les humains dans des contextes tels que l'Industrie 5.0, la santé ou les environnements domestiques. Cependant, les interactions physiques homme-robot (pHRI) sûres et naturelles restent difficiles en raison de la forte variabilité des comportements humains. Cette thèse vise à développer un cadre d'interaction physique homme-robot personnalisée et proactive, permettant au robot d'anticiper et de s'adapter en temps réel au comportement de chaque individu. Les transferts d'objets et le transport de charges constitueront les principaux cas d'étude, avec une validation sur des bras robotiques et des robots humanoïdes.

L'hypothèse centrale est que l'analyse et la prédiction en temps réel du comportement humain sont indispensables pour assurer des interactions robustes, sûres et personnalisées. Le projet cherchera ainsi à identifier les stratégies biomécaniques et comportementales mises en oeuvre lors de tâches collaboratives, afin de les intégrer à la génération des mouvements du robot, avec une attention particulière portée à la coordination spatiale et temporelle entre les partenaires.

La méthodologie combine biomécanique, analyse du mouvement, vision par ordinateur, robotique, contrôle optimal et apprentissage profond. Dans un premier temps, la collaboration humain-humain sera étudiée à l'aide d'un système de capture du mouvement tridimensionnelle et de capteurs d'effort six axes fixés sur l'objet manipulé. Ces mesures fourniront des données cinématiques et dynamiques synchronisées permettant l'estimation de la cinématique, des forces et des couples articulaires du membre supérieur. L'analyse des données permettra l'identification des mécanismes de coordination, des invariants mécaniques et des régularités de comportement exploitables par un robot. Dans un deuxième temps, ces données serviront à développer un cadre de contrôle optimal inverse (IOC) combiné à l'apprentissage profond permettant d'identifier une fonction de coût décrivant les stratégies de mouvement humain. Ce modèle prédira le mouvement à court terme à partir de la pose actuelle de l'humain et de l'historique récent. Une approche par fenêtre glissante intégrera progressivement les informations propres à chaque individu, permettant au modèle de s'adapter de manière personnalisée à chaque sujet.

Enfin, la perception de l'humain, la prédiction du mouvement et le contrôle robotique seront intégrés dans une architecture unifiée. Un système de vision déjà développé par les partenaires fournira la perception en temps réel de la posture humaine, permettant au robot d'anticiper les mouvements futurs et d'adapter proactivement sa trajectoire. L'approche sera d'abord validée avec des bras robotiques, puis étendue à des robots humanoïdes.

Le principal résultat attendu de cette thèse est un système intégré permettant une interaction physique souple, sécurisée et personnalisée avec un système robotique. La thèse permettra de mieux comprendre les mécanismes biomécaniques et comportementaux de la collaboration humain-hmain, notamment la variabilité inter- et intra-individuelle et les stratégies de coordination. Elle conduira également au développement de modèles IOC et d'apprentissage profond personnalisés capables de prédire le comportement humain à partir d'observations en temps réel. Le cadre proposé permettra ainsi au robot d'adopter un comportement proactif plutôt que purement réactif.

À terme, ce projet vise à proposer un cadre transférable à différentes plateformes robotiques, notamment des manipulateurs fixes ou mobiles et des robots humanoïdes. Les algorithmes et bases de données produits pourront contribuer au développement de robots collaboratifs plus sûrs et naturels et au déploiement d'interactions physiques homme-robot personnalisées dans les environnements de l'Industrie 5.0.
Humans are soft, flexible and their behaviour has been optimised through learning mechanisms and through evolution. Despite significant advancements in robotics, current technology has yet to enable robots to safely and seamlessly engage in collaborative handover and load-carrying tasks with humans. The precise coordination mechanisms facilitating such collaboration between humans remain elusive, even though humans are adept at developing intricate and resilient behaviours during interactions.

Moreover, investigations within the realm of Social and Human Sciences have explored the acceptance of collaborative robots (cobots) in the context of Industry 5.0, suggesting various adaptations within factory settings to enhance their integration into industrial domains (Prassidia et al., 2022; Liao et al., 2023).

Meanwhile, the domain of Physical Human-Robot Interaction (pHRI) has evolved into a highly dynamic research area focused on controlling robots in accordance with human behaviour. To tackle the challenge of evolving human behaviours, numerous research teams have introduced innovative control strategies for robots. One prominent approach for facilitating gentle interactions during load-carrying tasks is variable admittance control (Mujica et al., 2023), which has already demonstrated the feasibility of robust co-manipulation with lightweight loads. To recognise human intention and attention in a collaborative task, Wong et al. (2023) proposed a multimodal strategy based on supervised machine learning trained on touch location, human pose, and gaze direction. In a different context, researchers have utilised haptic feedback from human partners to dynamically adjust the kinematics of robots employing variable admittance control. These strategies enhance collaborative robot performance with integrated safety features. In terms of handover tasks, Pan et al. (2019) have controlled robot kinematics and utilised Bézier curves to achieve smooth, minimum-jerk motions, thereby enabling rapid reactions and minimal downtime for humans between object transfers.

The integration of Artificial Intelligence (AI) and advanced deep-learning models to facilitate human-friendly handovers was initially explored by Kupcsik et al. (2018). Preliminary experiments demonstrate that robots can learn to hand over objects naturally and adapt to human motion dynamics. More recently, Christen et al. (2023) have utilised similar frameworks based on trained robot agents in simulated environments to develop vision-based robots capable of performing human-to-robot handovers.

These existing studies have not concurrently addressed various subtasks within the same framework, whereas our objective is to facilitate seamless handover and load-carrying interactions between humans and robots. However, to achieve continuous and safe handover and load-carrying tasks, significant foundational innovations are necessary. Previous research has lacked integration of a human model to enhance robot proactivity and adaptability during physical interactions. Indeed, the absence of a musculoskeletal system model has limited these developments to quasi-static transfers, requiring humans to maintain static poses to mitigate the risk of collisions with the robots and potential injuries.

Notably, the partners involved in the project have already developed proactive motion generation for a humanoid robot interacting with humans based on human modelling (Maroger et al., 2022). Accordingly, this thesis aims to transfer these methods to real robotic platforms during object handover tasks, which require precise temporal and spatial coordination to ensure human safety.
Today, robots are increasingly regarded as viable solutions for assisting humans in various domains such as Industry 5.0, healthcare facilities, and household settings. However, ensuring direct and safe interactions with humans remains a formidable challenge due to the large variability of human behaviour. Acceptance of robots in industry, however, requires the construction of an anthropomorphic motion that can be understood by the subject interacting with the robot. In this context, the primary aim of this PhD. project is to advance the generalisation of personalised physical human-robot interaction (pHRI) by establishing a comprehensive framework for collaborative engagements. Central to this objective is the hypothesis that, given the substantial variability in human dynamics, real-time analysis of human behaviour is imperative to tailor interactions, ensuring robustness and safety. This necessitates a proactive approach by the robot to effectively improve the interactions. A standard scenario involving handover and load-carrying tasks will serve as a final testbed, with humanoid robot (Kawasaki Heavy Industries Kaleido) or arm manipulators, providing invaluable support for this case study. The ultimate framework, whose proof-of-concept will be demonstrated within the project, aspires to be deployable in industrial settings for long-term use and will undergo testing across various robotic platforms.

To realise this ambition, this Ph.D. project will tackle scientific challenges related to mechanically observing human-human interactions during handovers and load-carrying tasks. The concept of handover embodies a collaborative endeavour between two entities, characterised by spatial and temporal coordination. Consequently, the primary obstacle confronting the robot pertains to achieving precise positioning within the requisite temporal framework. This involves extracting key dynamic criteria, from humans, essential for integration into a customised inverse optimal control framework to accurately generate human-robot interaction behaviours. Personalising robot motion generation will be facilitated through real-time estimation of the human's pose interacting with the robot. Within this context, the team already developed (d'Haene et al., 2026) an embedded vision-based system capable of real-time localisation of human states, thereby enhancing the proactive motion of the robot. Indeed, knowing the past and present state of the human subject, we hypothesise in this Ph.D. project that the robot is capable of estimating the intention and predicting the future behaviour of the human subject using the optimal control and deep learning frameworks, and adapting accordingly in real time and in a proactive and personalised way.

In this framework, the final aim of this Ph.D. project is to deliver a complete system that integrates robot, vision-based system, human model and algorithms to generate soft and secure human-robot interactions.
This Ph.D. proposes an interdisciplinary methodology combining biomechanics, human movement analysis, computer vision, robotics, optimal control and deep learning methodologies to develop personalised and proactive physical human-robot interaction during object handover and load-carrying tasks. The methodological framework is organised into three interconnected steps. The first stage of the thesis consists of a biomechanical analysis of human-human collaboration, in which several pairs of participants will perform handover and load-carrying tasks under different experimental conditions. The experimental setup will combine a large-volume 3D Vicon motion-capture system with a six-axis force sensors attached to the shared object, enabling the measurement of object and interaction forces. The resulting recordings will provide simultaneous kinematic and kinetic information on both members of each dyad. Human movement will subsequently be represented using musculoskeletal models implemented through customised Python and OpenSim tools. Inverse kinematics will be used to reconstruct the three-dimensional motion of relevant body segments, while inverse dynamics will combine these kinematics with the forces measured on the object to estimate the forces and torques acting at the wrist, elbow and shoulder. Particular attention will be paid to the synchronisation between the two participants, the relative dynamics of their end-effectors and mechanical invariants. Machine-learning and data-analysis approaches will additionally be used to identify relevant patterns in the experimental database.

In parallel, the human movement data obtained in the first work package will be used to construct an Inverse Optimal Control (IOC) framework (Ishigaki et al., 2026) reinforced by deep learning analysis. Mean human trajectories will initially be used to estimate an optimal cost function describing the underlying movement strategy during handover, load carrying or other motions involving contacts (Ogura et al., 2026; Miyake et al., 2024; Wang et al., 2024). The resulting Optimal Control model coupled with deep learning algorithm will then predict the near-future human hand trajectory from the current pose and recent movement history. A sliding-window strategy will progressively replace population-average information with the individual's current and recent state, thereby enabling personalised prediction. Finally, these perception and prediction modules will be integrated into robot-control architectures and validated experimentally. Initial integration will use arm manipulators equiped with a 3D vision-based system developed by the partners (d'Haene et al., 2026), to detect real-time the pose of the humans in interaction with the robots. The final validation will combine handover and load-carrying into a continuous collaborative task and will extend the approach to humanoid robot.

Les candidats à cette thèse doivent être titulaires d'un master ou d'un diplôme d'ingénieur en robotique et/ou en (bio)mécanique. Ils doivent posséder des compétences en programmation Python. Une connaissance des méthodologies d'intelligence artificielle appliquées à la robotique et un intérêt pour l'expérimentation constitueraient un atout. La maîtrise de l'anglais est indispensable.

Application link : https://edd-projets.utoulouse.fr/