Léo Viguié

End-of-studies internship at ArianeGroup in the GNC (Guidance, Navigation, and Control) field, as part of the hypersonic re-entry program.

Development of reinforcement learning algorithms to optimize guidance and control laws during the atmospheric re-entry phase.

Work including:

• Modeling of launcher dynamics for the re-entry phase,

• Implementation of the learning environment and training of agents in Python (TensorFlow / PyTorch), then porting and integration of algorithms into the Matlab environment,

• Performance optimization (robustness, accuracy, energy consumption, compliance with physical constraints).

This project allowed me to combine skills in dynamic modeling, optimal control, and AI applied to aerospace.

Project carried out within the startup Differs during my training at CentraleSupélec.

Responsible for processing consumption data (choice of relevant granularity, product scale) and developing time series models (ARIMA, Prophet, DeepAR) to predict the evolution of sales for medium-sized stores.

Comparison of models and evaluation of performance using metrics such as MAPE and RMSE to optimize forecast accuracy.

Development of an intelligent controller for a simulated autonomous vehicle, using Deep Reinforcement Learning. Training of an agent (DDPG) via PyTorch to adjust speed and direction to follow a predefined trajectory stably and reactively. The simulation environment was integrated under ROS with virtual sensors. The project focused exclusively on control (low-level steering), without perception or planning.