Phd Position F/m Phd On Modular Model Predictive Control Architecture for Human-robot Collaboration

Le descriptif de l’offre ci-dessous est en Anglais_

**Type de contrat**: CDD

**Niveau de diplôme exigé**: Bac + 5 ou équivalent

**Fonction**: Doctorant

**Niveau d'expérience souhaité**: Jeune diplômé

**A propos du centre ou de la direction fonctionnelle**:
The Inria center at the University of Bordeaux is one of the nine Inria centers in France and has about twenty research teams.. The Inria centre is a major and recognized player in the field of digital sciences. It is at the heart of a rich R&D and innovation ecosystem: highly innovative SMEs, large industrial groups, competitiveness clusters, research and higher education players, laboratories of excellence, technological research institute...

**Contexte et atouts du poste**:
This PhD thesis will take place within the Auctus team at the Inria center at the University of Bordeaux.

The Inria center at the University of Bordeaux is one of the nine Inria centers in France and has about twenty research teams. The Inria centre is a major and recognized player in the field of digital sciences. It is at the heart of a rich R&D and innovation ecosystem: highly innovative SMEs, large industrial groups, competitiveness clusters, research and higher education players, laboratories of excellence, technological research institute, etc.
- Vincent Padois, senior research scientist at Inria, whose core scientific expertise focuses on robot control, especially in the collaborative context;
- Sébastien Kleff, postdoctoral scholar at Inria, whose core scientifc expertise lies in Model Predictive Control for robots in physical interaction.

**Mission confiée**:
While the first industrial robots were conceived to relieve humans from strenuous and hazardous tasks, robots are now primarily seen as a means to enhance productivity - both in mass production environments and in settings that demand industrial agility. Nevertheless, humans remain, for both organizational and societal reasons, the cornerstone of industry at large. Fully replacing them is still far from being technically feasible - despite the media hype following recent advancements in artificial intelligence - and remains highly debatable from a societal model perspective. It is therefore of paramount importance to continue improving workers’ health and working conditions. This is not only an ethical imperative but also an economic necessity [1], particularly in the context of a rapidly aging European workforce.

The growing ability of robots to produce both advanced motion and complex physical interaction - enabled by progress in actuation and sensing - makes them increasingly attractive as assistive systems. Yet, their capacity to effectively support humans in physically demanding tasks necessitates close human-robot collaboration (HRC), which raises several challenges. In particular, effective assistance requires the robot to adapt its behavior to the dynamic context of collaboration, such as human movement or environmental changes. Pre-planned motions or finely tuned controllers tailored to specific use-cases often lead to impasses. Recently, however, modern control techniques based on numerical optimization - such as nonlinear Model Predictive Control (MPC) [2] - have demonstrated strong potential for automating the synthesis and online execution of complex motions.

The proposed PhD thesis aims at addressing this issue through a modular control architecture centered on MPC. It is hypothesized that such an architecture is essential for equipping assistive robots with a general control framework capable of producing reactive yet optimal control decisions in uncertain environments.

Indeed, MPC and more particularly nonlinear MPC has shown great successes in synthesizing online complex motions on various robots ranging from industrial manipulators [3, 4] to quadrupeds [5, 6, 7] and humanoids [8, 9, 10]. The main advantage of MPC lies in its ability to re-optimize on the fly the robot’s trajectories based on a predictive model and physical constraints. However constructing accurate predictive models of the human and the shared environment is notoriously difficult [11, 12, 13].

Concurrently, learning methods (reinforcement learning [14], learning from demonstration [15]) have led to remarkable results on real hardware thanks to their ability to learn control policies mapping high-dimensional observations to low level robot actions [16] or latent representations [17] (e.g. human model). While they allow to push robots robustly to their performance limits in complex tasks, learned policies are often task-specific and lack generalization: transferring them across different tasks (e.g. to new control objectives or physical settings) is not trivial and usually requires to re-train from scratch. Lastly, they do not naturally allow to specify operating constraints, which is critical in HRC.

In such a framework, higher-level semantic planning can feed s-MPC with a list of, context dependent, desirable control obje