Fiche de CHENU Alexandre

Titre : Autonomous learning of complex robotics behaviors from a single demonstration.
Date : 10 mars 2023
Heure: 14:00
Résumé : This thesis evaluates the applicability of Deep Reinforcement Learning and Neuro-Evolution for learning neural control policies in robotics. After highlighting the limitations of Neuro-Evolution, we present a new Deep Reinforcement Learning approach. This approach uses a single demonstration to learn a goal-conditioned policy (GCP). This GCP is trained to control the robots between successive low-dimensional goals that are extracted from the demonstration. The robot reproduces the demonstrated complex behavior by achieving those successive goals. For example, in a humanoid robot locomotion task, the GCP can be used to control the robot between successive humanoid torso positions which results in walking behavior. We propose several variants of this approach and apply it to different robotic tasks, including grasping and locomotion tasks.
Lieu : Sorbonne Université Campus Pierre et Marie Curie Tour 66 Salle 310