JKTDrive: Early Explorations of Visual-Language Model on Autonomous Driving in Indonesia

Abstract

As existing driving datasets predominantly focus on urban environments in developed regions, their effectiveness in complex traffic conditions typical of developing regions remains uncertain. To address this gap, we introduce JKTDrive, a multiple-choice question (MCQ) benchmark designed to evaluate vision-language models (VLMs) in challenging urban driving contexts. JKTDrive comprises 273 image-question pairs across three categories: Perception and Reasoning, Safety Compliance, and Urban Flow Complexity. We benchmark five VLMs, with InternVL2.5 achieving the highest accuracy (81.32%), followed by GPT-5 (80.22%), Qwen2.5-VL (75.82%), Ovis1.6-Gemma2.5-VL (71.79%), and LLaVA-1.5 (47.62%). We further analyze performance across categories to identify model-specific strengths and weaknesses. The results demonstrate both the potential of VLMs and their remaining limitations in handling complex, real-world driving scenarios in developing regions. The dataset is publicly available at https://github.com/ChronoStellar/JktDriveVLM.