IMPACT-Agent: Intelligent Multimodal Prompt-Driven Agent for Collaborative Teaming

초록

The future of disaster response, search and rescue (SAR) technology is rapidly evolving with the integration of Artificial Intelligence (AI) and intelligent unmanned platforms. Manned-Unmanned Teaming (MUM-T) is revolutionizing on-site disaster structure and emergency rescue strategies by synchronizing manned and unmanned vehicles to support operations. This paper investigates the potential applications of Multimodal Large Language Models (MLLMs) in intelligent MUM-T systems. We present a system capable of using aerial imagery to identify and locate survivors and hazards/risks, and to generate safe path plans for unmanned ground vehicles (UGVs). The main contributions are: 1) the implementation of a Few-Shot In-Context Learning (ICL) method for MLLMs, which eliminates the need for fine-tuning or re-training the AI model, and 2) the application of the Chain-of-Thought (CoT) prompting technique. The ICL method reduces the time required for data gathering and model training, while the CoT reasoning technique improves the path planning generated by MLLMs. Our experiments, using aerial imagery, demonstrate that the proposed system achieves up to 86% mean Average Precision (mAP) in survivor and hazard identification through 5-shot ICL. Furthermore, applying the CoT technique doubled the UGV's path generation success rate from 20% to 40%. The proposed MLLM-based system effectively performed open-vocabulary object detection without additional training. This research highlights the potential of MLLMs for diverse and complex disaster relief missions in intelligent MUM-T systems, presenting an innovative approach to enhance capabilities and mission success while reducing human risk and operational costs. A demonstration video of our system is available at https://youtu.be/vReaSMDWu8o?si=RluQsqS4O_j5lVcS

키워드