EdgeCurator: Efficient Cross-Modal Grounding for Real-Time Language-Guided UAV Navigation

Abstract

Deploying cross-modal representations on a micro Unmanned Aerial Vehicle (UAV) requires accurate language–vision alignment and low-latency inference under tight compute constraints. Heavy Vision-Language Models (VLMs) provide strong semantic reasoning but are often unsuitable for an onboard control loop, whereas lightweight visual backbones commonly defer language fusion to a separate planning stage. We present EdgeCurator, a compact end-to-end early-fusion framework for real-time language-guided UAV navigation in access-controlled indoor spaces. Its Semantic-Spatial Grounding (SSG) Head converts lightweight text embeddings into semantic queries and uses conditioned deformable cross-attention to sparsely sample multi-scale visual features. On the custom Expo-VLN benchmark, EdgeCurator uses 14.8M parameters, runs at 54 FPS on an NVIDIA Jetson Orin NX, and reaches 67.3% grounding accuracy and 62.8% navigation success. We additionally report 60 controlled physical trials with two target types, achieving 78.3% success. These trials test onboard Sim2Real feasibility for pre-opening inspection and curation assistance; they do not establish operation around visitors or generalization to full exhibition deployments. The results position EdgeCurator as an efficiency-oriented cross-modal representation and grounding system for resource-constrained aerial agents.