LatentAM: Real-Time Large-Scale Robotic Semantic Mapping via Online Dictionary Learning

rsasaki0109 · x · 2026-08-15

LatentAM is an online 3D Gaussian Splatting (3DGS) mapping framework for open-vocabulary robotic perception. It builds scalable latent feature maps via online dictionary learning, being model-agnostic and pretraining-free. Each Gaussian primitive has a compact query vector converted to approximate VLM embeddings via attention with a learnable dictionary. Voxel hashing manages maps, with GPU optimizing active local map and CPU storing global map. Experiments show superior feature reconstruction fidelity and near-real-time speed (12-35 FPS).

Original post →

More from Embodied

Embodied channel →