Συντάχθηκε 14-08-2024 16:39
Τόπος: Λ - Κτίριο Επιστημών/ΗΜΜΥ, 145Π-58
Σύνδεσμος τηλεδιάσκεψης
Έναρξη: 20/08/2024 13:00
Λήξη: 20/08/2024 14:00
Abstract
In the realm of advanced robotics, enabling machines to perceive and interact with their environments as seamlessly as humans remains a significant challenge. This talk will explore the advancements in self-supervised representation learning, visuotactile fusion, and robot learning for robotic manipulation. Central to this exploration will be the development of robust, generalizable, and efficient representations from unlabeled data, leveraging techniques such as autoencoding, contrastive learning, and regularization-based methods.
In robotic manipulation, integrating visual and tactile sensory data is crucial for performing complex, nuanced tasks that require a detailed understanding of the environment. Traditional robot learning models often rely heavily, or solely, on visual data, which can be limited in certain contexts, such as when dealing with occlusions. Tactile data, on the other hand, provides essential and more local information about texture, force, and compliance, which is vital for precise manipulation but is often underutilized.
The talk will discuss novel frameworks that efficiently combine visual and tactile sensory data, leading to improved material property classification and more successful grasping outcomes, thereby enhancing manipulation capabilities. Additionally, new models will increase the sample efficiency and robustness of reinforcement learning algorithms, resulting in more efficient and reliable decision-making.
Looking ahead, the talk will conclude with future directions, including overcoming the challenges of Sim2Real transfer and advancing imitation learning techniques, to further enhance robotic performance in real-world applications.
Keywords: Self-supervised representation learning, Visuotactile fusion, Reinforcement learning, Imitation Learning
About the Speaker
Fotios has been a doctoral student and university assistant at the Chair of Cyber-Physical Systems (CPS) since March 2022. His research focuses on representation and robot learning, employing self-supervised learning methods, both contrastive and non-contrastive, as well as reinforcement learning techniques, specifically targeting manipulation tasks.
Fotios received a Diploma in Electrical and Computer Engineering (equivalent to an Integrated Master in Engineering) from the Technical University of Crete (Greece) in 2019. Before starting his PhD at the University of Leoben, Fotios held positions as a teaching assistant at the University of Texas at Arlington (USA), a research assistant at the National Center for Scientific Research Demokritos in Athens (Greece), and a research intern at Toshiba Research Europe in Cambridge (UK).
His work includes contributions to representation learning, reinforcement learning for robotic manipulation, healthcare robotics, and dialogue systems. He is also active in teaching, technical skill development, outreach, reviewing, and conference activities.
For more information, please refer to his personal webpage www.lygerakis.com
Σύνδεσμος εκδήλωσης
