Researchers' Co-GLANCE boosts robot team accuracy 36% with 350x faster perception

Perceptual uncertainty is a critical challenge for heterogeneous robot teams operating in unstructured outdoor environments, where occlusions and limited viewpoints degrade scene understanding. Existing solutions rely on cloud-based vision-language models (VLMs) that are computationally prohibitive for onboard inference and lack calibrated uncertainty quantification. Researchers from the University of Texas at Austin propose Co-GLANCE, a distributed perception and decision-making system that distills VLM semantic reasoning into a lightweight end-to-end model for occlusion segmentation and robot allocation, eliminating cloud dependency.

Co-GLANCE combines conformal prediction with selective abstention to provide statistically valid coverage guarantees for segmentation, allocation, and detection outputs. These calibrated uncertainty estimates trigger active perception, dispatching the most appropriate robot to acquire informative viewpoints. Across real-world scenarios, Co-GLANCE achieves 25% higher occlusion segmentation accuracy and 36% higher robot allocation accuracy than cloud-based VLM baselines, while reducing per-frame inference latency by 350x. The team also releases an air-ground dataset to accelerate research in autonomous multi-robot systems.