Doctoral thesis defense in Computer Science by Ms. Hiba SEKKAT

The Euromed University of Fez (UEMF) is pleased to inform the public of the

defense of doctoral thesis in “Computer Science”

The thesis defense will take place on Saturday October 26, 2024 at 10:00 a.m. at the UEMF

Location : the Great Hall of the Incubator (LOC001994)

The thesis will be presented by Ms. Hiba SEKKAT

Under the theme :

“Integrated Approaches in Robotic Systems: Advancing Grasping with Deep Reinforcement Learning and Digital Twin Simulations for Domestic and Humanoid Robots”

Abstract 

In the fast-evolving field of robotics, grasping capabilities of domestic and humanoid robots are crucial for integration into daily life; this has been especially highlighted during the COVID-19 pandemic which highlighted the need for robots to assist with everyday tasks, when human contact was reduced, particularly for vulnerable populations such as the elderly and individuals with disabilities. Among the most important tasks to be performed by a robot are grasping tasks, enabling interaction with objects, and thus, the ability to pick up items, arrange places, and help with everyday chores. However, current robotic systems often find it hard to be adaptive and efficient when operating in natural environments, especially home-like environments, that are inherently dynamic and cluttered. Dealing with and overcoming these challenges would hence require innovation that enables dealing with the complexity and variability of realworld settings. This thesis thus takes up the critical challenges of grasping in robotic manipulation through the development of a high-fidelity digital twin simulation environment designed to facilitate the future integration of Deep Reinforcement Learning algorithms. Meanwhile, the development and testing of such algorithms using digital twin simulations would provide a highly accurate controlled environment, making sure the application on a real robot is well-prepared. Therein, the initial step involved a detailed review of state-of-the-art Reinforcement Learning algorithms to identify the most suitable one for robotic grasping tasks across various settings. From this review, we selected the Deep Deterministic Policy Gradient algorithm that is particularly well-suited for our work environment, leading to creating a novel grasping system with high accuracy compared to the benchmark methods. Building on these promising results, we developed an efficient simulation system—a high-fidelity digital twin of the Pepper robot created using the Robot Operating System 2 framework—that supports continuous improvement and helps progress the field. With most movements having a mean absolute error within a range of 0.01–0.02 radians, this digital twin—validated by comparing joint angles in both simulation and real-world data—lends a robust simulation platform for the eventual integration of intricate machine-learning algorithms . These results show how Deep Reinforcement Learning and digital twin technologies can be combined to overcome current limitations and open the door for more capable and independent home robots. This work is thus one of the competitive methodological steps toward better robotic grasping and sets the stage for extensions into autonomous robotics. All of these findings have important implications for developing capacities in socially interacting robots, notably the Pepper robot, in improving overall quality of life and offering substantial benefits for vulnerable populations.

Summary 

In the rapidly evolving field of robotics, the grasping capabilities of domestic and humanoid robots are crucial for integration into daily life; This was particularly highlighted during the COVID-19 pandemic, which highlighted the need for robots to help with everyday tasks, when human contact was reduced, particularly for vulnerable populations such as the elderly and people with reduced mobility. Among the most important tasks to be performed by a robot are grasping tasks, which allow it to interact with objects, and therefore pick them up, arrange places and help with everyday tasks. However, current robotic systems often struggle to adapt and be effective when operating in natural environments, particularly domestic environments, which are inherently dynamic and cluttered. Meeting and overcoming these challenges will therefore require innovations to cope with the complexity and variability of real-world environments. This thesis therefore addresses the critical challenges of grasping in robotic manipulation through the development of a high-fidelity digital twin simulation environment designed to facilitate the future integration of deep reinforcement learning algorithms. At the same time, developing and testing such algorithms using digital twin simulations would provide a highly precise controlled environment, ensuring that the application on a real robot is well prepared. The initial step consisted of a detailed review of the state of the art of reinforcement learning algorithms in order to identify the best suited for robotic grasping tasks in different contexts. From this study, we selected the Deep Deterministic Gradient Policy algorithm which is particularly well suited to our working environment, which allowed us to create a new gripping system with great precision compared to the methods of reference. Based on these promising results, we have developed an effective simulation system - a high-fidelity digital twin of the Pepper robot created using the "Robot Operating System 2" framework - which enables continuous improvement and contributes to advance the field. With most movements having an average absolute error between 0.01 and 0.02 radians, this digital twin - validated by comparing joint angles in simulation and real data - provides a robust simulation platform for eventual integration complex machine learning algorithms. These results show how deep reinforcement learning and digital twin technologies can be combined to overcome current limitations and pave the way for more capable and independent home robots. This work therefore constitutes one of the competitive methodological steps towards better robotic grasping and opens the way to extensions in the field of autonomous robotics. All of these findings have important implications for developing the capabilities of social interaction robots, including the Pepper robot, by improving overall quality of life and providing substantial benefits to vulnerable populations.

This thesis will be presented to the jury members:

Share