GPT-4off: On-Board Traversability Probability Estimation for Off-Road Driving via GPT Knowledge Distillation

초록

This paper proposes a framework for predicting traversability probability in off-road environments by distilling knowledge from large language models (LLMs) such as GPT-4o into lightweight models. The GPT-4off approach utilizes GPT-generated data to train a compact model capable of real-time operation on edge devices, such as the NVIDIA Orin board. Unlike traditional systems that focus on identifying traversable areas, this study emphasizes the prediction of traversability probability to facilitate faster decision-making in complex environments. This is particularly advantageous for unmanned ground vehicles, for which obstacles and terrain variability present significant challenges. The GPT-4off framework improves real-time performance through knowledge distillation and domain-specific optimization, ensuring efficient resource use while maintaining LLM-level performance. Experimental results on the RUGD off-road dataset show that the lightweight model achieves GPT-level performance while being deployable on edge devices. This framework effectively reduces human annotation costs and RAM power consumption, improves the practicality of off-road autonomous driving systems, and demonstrates the potential of leveraging LLM capabilities for low-power real-time applications.

키워드