Bilal Tout's defense (automatic department)

In recent years, physical human-robot interaction has become an important research topic, for example for rehabilitation applications. This thesis aims to improve these interactions, as part of the development of model-based controllers, through parametric identification approaches to the models of interacting systems. The aim is to develop identification methods taking into account the variability and complexity of the human body, and using only the sensors of the robotic system to avoid the addition of external sensors.

The various approaches presented in this thesis are experimentally tested with a one-degree-of-freedom (1-DDL) robotic system for interacting with a person's hand.

After a 1st chapter presenting the state of the art, the 2nd chapter addresses the identification methods developed in robotics as well as the filtering issue, analyzed in simulation and experimentally. The issue of low-pass filter tuning is addressed, and in particular the choice of cut-off frequency, which remains a delicate one for a non-linear system.

To overcome these difficulties, a filtering technique using an extended Kalman filter (EKF) is developed based on the dynamic model of the robot. The proposed EKF formulation allows tuning according to known sensor properties and confidence in the initial parameter estimate. This method is compared in simulation and then experimentally with various existing methods, analyzing sensitivity to filter initialization and tuning. The results show that the proposed method is promising if the EKF is properly tuned.

Chapter 3 focuses on the continuous identification of dynamic model parameters of a passive system interacting with a robotic system, by combining payload identification methods with online algorithms, without external sensors. These methods are validated in simulation and experimentally using the 1-DDL system, whose handle is attached to elastic bands to mimic a passive human joint. Analysis of the effect of fitting the methods online highlights that a compromise is needed between convergence speed and the accuracy of parameter estimates. Finally, comparison of payload identification methods shows that methods separately identifying the parameters of the robotic system and the passive human give better accuracy and lower computational complexity.

The 4th chapter focuses on identification during human-robotic system interaction. A quadratic stiffness model is proposed to better represent the behavior of the passive human joint than a linear model. Subsequently, this model is used with an iterative identification method based on outlier rejection, to detect human muscle activity without external sensors. This method is compared experimentally with a non-iterative method using electromyography (EMG) signals, by adapting the system to 1-DDL to interact with the wrist and enable the flexor and extensor muscle activity of two subjects to be assessed. The proposed iterative method without EMG signals produces results close to those obtained with the method using EMG signals, when a model is chosen that accurately represents the behavior of the passive human joint. The muscle activity detection results obtained with these two methods show a satisfactory level of similarity with those obtained directly from EMG signals.

Rapporteurs:

• JANOT Alexandre, Research engineer, ONERA Palaiseau.
• LAROCHE Edouard, University Professor, University of Strasbourg, Icube Laboratory.

Examiner:

• SIEGLER Isabelle, University Professor, Université Paris Saclay, CIAMS Laboratory.

Thesis supervisor:

• VERMEIREN Laurent, Professeur des Universités, Université Polytechnique Hauts-de-France, Laboratoire LAMIH (UMR 8201).

Co-supervisors:

• CHEVRIE JASON, Maitre de conférences, Université Polytechnique Hauts-de-France , Laboratoire LAMIH (UMR 8201).
• DEQUIDT Antoine, Maitre de conférences, , INSA Hauts-de-France, Laboratoire LAMIH (UMR 8201).