Reinforcement Learning (RL) is an approach in Artificial Intelligence where an agent learns to make decisions by interacting with its environment, aimed at maximizing a defined reward over time. Despite its promising results in various tasks, RL is often seen sparingly in critical systems such as healthcare or aviation, owing to its difficulty in guaranteeing safety and reliability under unknown circumstances, a challenge termed as the ‘reality gap’. This often stems from the lack of ability to generalize well to unseen situations and a gap between simulations and real-world applications.

Research directions

Our research project seeks to address these issues head-on, with a primary focus on enhancing RL’s generalization capabilities.  We intend to enable the RL agent to learn meaningful and relevant representations of its environment, thereby boosting its ability to make accurate decisions in new, unexplored contexts. By creating robust algorithms capable of transferring learned skills from simulations to real-world scenarios (Sim2Real), we aim to bridge the ‘reality gap’ that hinders RL’s application in critical domains. Furthermore, the project will also explore RL for Adaptive Stress Testing (AST), a method to identify the most likely paths leading to system failures, thus, improving the safety of critical systems. Through these targeted efforts, we hope to widen the application range of RL, ushering it from the digital confines of simulations into critical, real-world systems, where its potential can truly be realized.

Main Publications


« Look where you look! Saliency-guided Q-networks for generalization in visual Reinforcement Learning », David Bertoin, Adil Zouitine, Mehdi Zouitine, Emmanuel Rachelson, NeurIPS 2022


« Causal Reinforcement Learning using Observational and Interventional Data », Maxime Gasse, Damien Grasset, Guillaume Gaudron, Pierre-Yves Oudeyer, NeurIPS 2022 Workshop 


« Local Feature Swapping for Generalization in Reinforcement Learning », David Bertoin, Emmanuel Rachelson, ICLR 2022


« Autonomous drone interception with Deep Reinforcement Learning », David Bertoin, Adrien Gauffriau, Damien Grasset, Jayant Sen Gupta, IJACI-ECAI 2022 Workshop: 12th International Workshop on Agents in Traffic and Transportation