Work

GenAI-Nav: Large-Multimodal-Model-Assisted Autonomous Navigation and Mission Reasoning for UAVs

Year: 2026
Type: article
Abstract: Unlike conventional Unmanned Aerial Vehicle (UAV) systems with static operations, long-horizon understanding, multimodal fusion, and safe decision-making can no longer be handled by conventional control and Reinforcement Learning (RL) methods. This study proposes GenAI-Nav, a Large Language Model (LLM)-assisted UAV navigation framework based on Temporal Mission Memory (TMM), which learns to retain context; a Multimodal Fusion Attention Controller (MFAC) that dynamically fuses multimodal information; a Failure-Aware Reasoning Layer (FARL) for prediction modeling; and a Safety-Constrained LLM Action Filter (SCLAF) for safety constraint-based action filtering. These elements allow for a single-pipeline use of Generative Artificial Intelligence (GAI) for unified temporal reasoning, multimodal perception, failure prediction, and a safety-aware decision-making system. The framework is assessed on high-level scenarios of the UAV system, which include Global Positioning System (GPS)-denied navigation, high winds, multi-obstacle, and disaster response missions. Results show that it consistently outperforms classical control, RL, and LLM baselines in various metrics including Mission Success Rate (MSR), Decision Safety Score (DSS), Temporal Consistency Index (TCI), Multimodal Alignment Score (MAS), and Failure Prediction Accuracy (FPA), showing improved robustness, safety, and reliability on complex UAV missions.
Language: English
Open Access status: diamond
OpenAlex ID: W7225164378
FWCI:
Cites: 21
Cited by:
Related to: 10