Description du sujet de thèse

Study of algorithms of reinforcement learning for robotics

Reinforcement learning (RL) has demonstrated significant potential for robotic
applications, particularly in legged locomotion. However, the efficiency of RL
solvers still falls short of practical requirements. Building on insights from
trajectory optimization, this project proposes using standard optimization
solvers as alternatives to popular RL algorithms like PPO or SAC. We will focus
initially on evolutionary strategies (ES), leveraging their competitiveness on
CPU architectures and exploring their scalability on GPU architectures for
substantial performance gains. Building on this foundation, we aim to refine
the problem formulation, including investigating multiple shooting methods to
enhance exploration in state space rather than control space. We will exhibit
the property and importance of this formulation, called transcription in
trajectory optimization and rarely consider in RL. Additionally, we will explore
incorporating gradients from the simulator or neural networks where feasible.
The ultimate goal is to apply these methods to real-world quadruped
locomotion tasks.

Contexte de travail

Team Gepetto

Le poste se situe dans un secteur relevant de la protection du potentiel scientifique et technique (PPST), et nécessite donc, conformément à la réglementation, que votre arrivée soit autorisée par l'autorité compétente du MESR.

Contraintes et risques

Nothing