DESIGN AND PRELIMINARY EVALUATION OF A HIERARCHICAL CONTROL AND LATENT PERCEPTION PIPELINE FOR UAV WAREHOUSE NAVIGATION

This thesis presents a modular deep reinforcement learning approach for the design and preliminary evaluation of a hierarchical control and latent perception pipeline for UAV (Unmanned Aerial Vehicle) warehouse navigation in NVIDIA IsaacLab. The core problem addressed is the design of a system capable of stable, collisionfree flight and accurate position tracking without relying on any pre-built map, semantic annotation, or global positioning signal, using only onboard depth sensing and proprioceptive state estimates. The proposed architecture decomposes the navigation problem into several components: a hierarchical controller, composed of a low-level velocity controller and a high-level position controller, trained via PPO (Proximal Policy Optimization) in a massively parallelized IsaacLab simulation; a compact latent representation of the raw depth image, produced by a 2D Variational Autoencoder trained offline on simulated depth frames, which compresses the collision-relevant information contained in the input image into a low-dimensional space; a second latent space, produced by a 3D deterministic autoencoder, that compresses the local occupancy map and the spatial visit score map around the drone; and a GRU (Gated Recurrent Unit) that allows the system to retain information about the environment over time. The overall objective was divided into two sub-problems. The first, simpler approach was obstacle-free and relied on a leaner training pipeline, yielding excellent results. The second, more complete approach was set in a more complex environment featuring obstacles and employed all of the components described above; it produced partial results and validated only some aspects of the overall problem.