Contact robots are increasingly used to assist humans in physical training and manufacturing tasks. However, their effectiveness is currently limited as control methods focus on system performance without considering the upcoming human user’s control. Here, we present a differential game-based controller for contact robots ensuring optimal interaction by predicting human motor control over their finite planning horizon. Using this model-predictive game (MPG) controller, we investigated human-robot co-adaptation in experiments, demonstrating that: (a) MPG interaction remains stable while reducing human effort; (b) the robot adapts to humans, identifying time-consistent individual interaction behaviors; (c) humans adapt to the robot, and their behavior can be modulated through an assistance meta-parameter adjusting the robot’s propensity to minimize human effort. These findings indicate that humans understand and adapt to a partner’s control strategy aligning with game theory principles. Furthermore, the assistance meta-parameter’s ability to guide humans toward specific interaction behaviors enables versatile robot-assisted systems for physical training and rehabilitation.

Contact robots are increasingly used to assist humans in physical training and manufacturing tasks. However, the effectiveness of current systems is limited as their control focuses on the system performance without explicitly considering the upcoming human user’s control. Here we present a differential game-based controller for contact robots that ensures optimal interaction with the human user by predicting their control during movement while considering their inherently finite planning horizon. Using this model-predictive game (MPG) controller, we systematically investigated human-robot co-adaptation in experiments, demonstrating that: (a) interaction with MPG remains stable for all participants while effectively reducing human effort; (b) the robot adapts to human behavior, identifying and characterizing individual motor control strategies that remain consistent over time and may be used as control biomarkers; (c) the human adapts to the robot’s behavior, and their interaction behavior can be modulated through an assistance meta-parameter. These findings indicate that humans can understand and adapt to a partner’s control strategy during movement, thereby exhibiting behavior consistent with game theory principles. Furthermore, the ability of the assistance meta-parameter to guide human users toward specific interaction behaviors can be used to develop versatile robot-assisted learning systems for physical training and rehabilitation.

Exoskeletons are appealing robotic devices to physically assist humans in various motor tasks. To provide a part of the effort required for performing a task, they should be intuitive to use and adapt to the user's goal. Differential game theory offers an interesting framework to formalize the shared control problem underlying physical human-robot interaction. In the present paper, we introduce an approach based on finitehorizon inverse differential games, which allows to iteratively infer the user's internal goals and design a Nash-equilibrium control policy. Here, we focus on a case study in which a user has to move a load along a target trajectory while assisted by a wrist exoskeleton. The method is first validated in simulations and then applied to the control of the HRX-1 wrist interface. The results show that the controller has a positive impact tracking performance and reduces the joint torque provided by the users while enabling them to remain active. Interestingly, it also yields to a more balanced sharing of task efforts and a better coordination between the robot and the user compared to a user-agnostic linear-quadratic control guidance.

Commande prédictive par jeux différentiels pour l'assistance intuitive du mouvement lors d'une interaction humain-robot: Les manipulateurs robotiques et les exosquelettes offrent des perspectives prometteuses pour des applications allant de la rééducation motrice à l'entraînement physique et aux tâches industrielles. Cependant, leur efficacité reste limitée, principalement en raison de la difficulté à anticiper les politiques de contrôle utilisées par l'utilisateur humain. La présente thèse explore l'hypothèse selon laquelle la modélisation de l'humain et du robot en tant qu'agents adaptatifs interagissant selon des principes de théorie des jeux pourrait améliorer leur interactions physique. Pour tester cette hypothèse, nous avons développé un cadre de travail basé sur le contrôle par jeu prédictif (MPG) dérivé de la théorie des jeux différentiels. L'objectif du contrôleur est d'anticiper la stratégie motrice humaine en estimant une fonction de coût représentative et en adaptant l'assistance robotique en conséquence. Cette approche a d'abord été validée par des simulations, puis expérimentalement avec un exosquelette de poignet pour des tâches de suivi de trajectoire. Les résultats de ces études ont démontré que le contrôle MPG (i) réduit l'effort humain tout en maintenant la stabilité de la tâche, (ii) capture les caractéristiques individuelles de la stratégie motrice qui restent cohérentes dans le temps, et (iii) permet de moduler les stratégies motrices humaines grâce à un méta-paramètre d'assistance régulant la minimisation de l'effort humain. Les résultats indiquent que, dans le tâches investiguées, les humains comprennent et s'adaptent à la stratégie de contrôle de leur partenaire en se conformant aux principes de la théorie des jeux. De plus, la capacité du méta-paramètre d'assistance à guider les humains vers des comportements d'interaction spécifiques permet de créer des systèmes polyvalents assistés par robot pour l'entraînement physique et la rééducation.