Goal-driven 3D Map Generation with Reinforcement Learning
An RL agent that dynamically allocates resolution in a 3D map under a tight memory budget.
This page is still a work in progress — for now it’s just a quick summary of the project.
Computer Vision and Geometry Lab, ETH Zürich (Daniel Barath) — Spring 2026
Trained an RL agent using PPO to dynamically allocate resolution in a 3D map, learning to keep high detail in regions relevant to the task at hand while compressing the rest, under a tight memory budget.
Built a Python library for 3D maps with online adaptive resolution (incremental delta-buffer updates, raycasting-based false-obstacle pruning) and integrated it with the PEANUT navigation agent in the AI Habitat simulator, cutting memory usage by 88.6% in test episodes while maintaining task success.