Abstract

Robust self-localization is a critical capability for autonomous humanoid robots competing in dynamic environments. We present ViTHL, a hybrid localization architecture that combines the representational power of vision transformers with geometric constraints for accurate and reliable robot pose estimation. Our approach leverages pre-trained ViT models for feature extraction from onboard camera feeds, fused with traditional odometry signals through a probabilistic filter. Evaluated on humanoid robots in RoboCup competition scenarios, ViTHL demonstrates improved localization accuracy over baseline methods under challenging lighting and field conditions.