ED Sciences Physiques et de l'Ingénieur
Shared-autonomy control for improving Human-Robot collaboration in haptic teleoperation
by Elio JABBOUR (Institut national de recherche en informatique et en automatique - Bordeaux - Sud-Ouest)
The defense will take place at 9h30 - Amphi 0108 ENSC 109 Avenue Roul 33400 Talence
in front of the jury composed of
- Vincent PADOIS - Directeur de recherche - Inria - Directeur de these
- Margot VULLIEZ - Chargée de recherche - Inria - CoDirecteur de these
- Paolo ROBUFFO GIORDANO - Directeur de recherche - IRISA/Inria- Université de Rennes - Examinateur
- Caroline VIENNE - Ingénieure de recherche - Laboratoire de Robotique Interactive du CEA6List - Examinateur
- Vincent CREUZE - Professeur des universités - Campus Saint Priest- Université de Montpellier - Rapporteur
- Mourad BENOUSSAAD - Maître de conférences - Université de Technologie Tarbes Occitanie Pyrénées- UTTOP - Rapporteur
Shared control frameworks assist human operators by blending their commands with autonomous, goal-oriented trajectories. However, conventional blending techniques often fail to guarantee the feasibility of the resulting motion or the optimality of the combined decision. This thesis addresses two principal gaps in shared control: 1) the lack of a blending arbitrator that unifies predictive foresight with verifiable safety in a computationally tractable manner, and 2) the flawed assumption that the autonomous assistance is correct, which leads to performance degradation and user-robot conflict when the system's world model is misaligned with reality. This work presents a control architecture that resolves both challenges. First, to address the arbitration gap, we formulate blending as a constrained optimal control problem. A Model Predictive Control for Blending (MPC-B) framework is proposed to compute a feasible blended trajectory via receding-horizon optimization, ensuring proactive compliance with all system and task constraints. Second, to address the assistance gap, we introduce a Dual-Component Adaptive Assistance Framework that corrects for model inaccuracies by treating the operator's input as a corrective measurement. This framework integrates a real-time Adaptive Kalman Filter to compensate for local, transient errors and an online N-Point Procrustes Analysis module to learn and correct for global, systematic misalignments over time. The proposed architecture was evaluated in two human-in-the-loop teleoperation studies. The experimental results demonstrate the superiority of the proposed frameworks compared to conventional blending and unassisted teleoperation. The MPC-B controller significantly improved safety by eliminating kinematic constraint violations, while the adaptive assistance framework successfully overcame significant model errors to improve task efficiency beyond unassisted capabilities. Taken together, the results validate the integrated architecture as a solution for safer and more effective human-robot collaboration, yielding quantifiable improvements in task performance, interaction quality, and operator workload.