Task-Driven Map Compression for Collaborative Exploration of Ground-Aerial Robot Teams

Sammanfattning

Collaborative multi-robot exploration relies on robots sharing spatial information to coordinate navigation and ensure safe deployments. However, in communication-restricted environments, such as subterranean environments, transmitting raw sensor data becomes unfeasible for real-time applications. To address this, this thesis proposes a direct map compression framework that compresses local submaps into latent representations using a Variational Autoencoder. Local submaps are represented as binary voxel grids of fixed size, with occupied voxels set to 1 and unoccupied voxels to 0. The pipeline uses a distance-based extraction strategy to capture local submaps along a robot's primary path. The framework's compression performance was compared with standard baselines, namely native OctoMap serialization and Google's Draco. The pipeline achieved compression ratios ranging from 7.48 to 11.38 while preserving the overall structure, enabling downstream use for path planning. Ultimately, this work shows that learning-based spatial representations can effectively overcome severe bandwidth limitations.