Overview
This project addresses the challenge of autonomous rover navigation in unstructured lunar environments. Developed during a traineeship at the European Space Agency (ESA), the goal was to evaluate whether Reinforcement Learning (RL) agents could outperform traditional deterministic planners in handling the unpredictability of the lunar surface
Massive Parallel Simulation: Isaac Lab
To overcome the limitations of traditional sequential simulators, I utilized NVIDIA Isaac Sim and Isaac LabThis allowed for the simulation of hundreds of rovers simultaneously on a single GPU, generating millions of transitions in hours. This massive scale was essential for training robust agents capable of generalizing across varied terrains and lighting conditions.
Advanced AI Methodologies
The core of the research involved a comparison of different RL paradigms:
- Recurrent PPO (LSTM): I demonstrated that memory is a functional requirement for space autonomy. Agents with recurrent layers successfully managed partial observability, "remembering" obstacles even when they left the sensor's field of view.
- World Models (TD-MPC2): I explored latent planning through the TD-MPC2 algorithm. By learning an internal predictive model of the world (TOLD architecture), the rover can "imagine" future trajectories and anticipate risks like slippage or steep slopes before they happen
Sim-to-Real Deployment
The entire pipeline—including a custom geometric perception system based on DBSCAN and RANSAC—was optimized for edge computing. It was successfully deployed on a Raspberry Pi 5 with a Hailo-8 AI Accelerator. Final validation took place at the LUNA Analog Facility (EAC Cologne), where the rover demonstrated robust navigation on real volcanic regolite, proving that agents trained in the metaverse can survive in the dust of the Moon.