Paul-Antoine Le Tolguenec

PhD · Deep Reinforcement Learning

Paul-Antoine
Le Tolguenec

PhD (CIFRE) at ISAE-Supaero with Airbus Commercial Aircraft, developing exploration methods for deep reinforcement learning applied to critical system testing. MSc in autonomous robotics from ENSTA. Passionate about RL and its applications to robotics.

GitHub LinkedIn Email
scroll

01 Projects

Crazyflie quadrotor hovering under a wind gust perturbation in the Genesis simulator
Deep RL Robustness Simulation Robotics

Genesis Robust Quadrotor

Robust-RL framework for drones on the Genesis simulator — 69 perturbations across 8 categories (physics, motors, sensors, wind, …), adversarial-training API (DR / RARL / RAP), Gymnasium-compatible.

GitHub
TrackMania IL policy driving the first track
Imitation Learning Offline RL Vision

TrackMania Imitation Learning

ACT behavioral cloning and offline IQL on TMNF, shared frozen DinoV2 encoder. Mac serves inference over TCP, Windows runs the game.

GitHub
Coming soon
RL · Multi-agent
Deep RL Multi-agent Simulation

Dogfight Drones

A quadrotor drone combat simulator where multiple autonomous agents fight each other using deep RL policies.

UAV · RL
Deep RL Robotics ROS

UAV Autonomous Docking

DDPG policy trained to dock a UAV autonomously. CNN + LSTM observations, deployed on Nvidia Jetson with ROS, validated in Gazebo simulation.

GitHub
Karst simulation Gazebo
Robotics Simulation ROS

Karst Robotics Simulation

Gazebo simulation for mobile robot navigation in underwater karst terrains, with an interval-analysis framework for uncertainty quantification.

GitHub

02 Blog and Seminars

Coming soon
Blog

Alternative to Robust RL

Exploring new directions beyond classical robust RL formulations for policy robustness under distribution shift and uncertainty.

Blog

Manifold Exploration

Smooth manifolds, Christoffel symbols, geodesics, and the geometry of curved spaces.

Read post →
Course

Introduction to Reinforcement Learning

A course on the fundamentals of reinforcement learning: MDPs, policy gradient, value functions, and modern deep RL methods.

View course →
Seminar

Diffusion Processes: a Formal Introduction

A rigorous treatment of diffusion processes: SDEs, score matching, and the mathematical foundations of modern generative models.

View slides →
Defense

PhD Defense: Exploration Methods for Deep RL

Slides from my PhD thesis defense on exploration methods for deep reinforcement learning applied to critical avionic system testing.

View slides →

03 Publications & Pre-prints

RAMP paper figure RAMP
Pre-print · 2024

Exploration by Running Away From the Past

Le Tolguenec, Wilson, Besse, Rachelson

Proposes RAMP, a method encouraging agents to explore by distancing themselves from past experiences, without reward shaping or intrinsic bonuses.

Paper
LEADS paper figure LEADS
NeurIPS 2024

Exploration by Learning Diverse Skills through Successor State Measures

Le Tolguenec, Wilson, Besse, Rachelson

Introduces LEADS, which creates a diverse set of skills to enhance exploration, without rewards or intrinsic bonuses, using successor state representations.

Paper
ISSTA paper figure ISSTA
ISSTA 2024 · ACM Distinguished Paper Award

Exploration-Driven Reinforcement Learning for Avionic System Fault Detection

Le Tolguenec, Rachelson, Besse, Wilson

Uses heuristic-based exploration to identify failure scenarios in avionics systems, improving flight system reliability through RL-driven robustness evaluation.

Paper
Curiosity-ES paper figure Curiosity-ES
ACM TELO · 2023

Curiosity Creates Diversity in Policy Search

Le Tolguenec, Rachelson, Besse, Wilson

Introduces Curiosity-ES, using curiosity as a fitness metric in evolutionary strategies to explore reward-sparse environments and discover diverse policies.

Paper