Building smarter maps with GPT-4o vision fine-tuning

Southeast Asia is an extremely challenging region for precise mapmaking due to narrow roads, motorcycle-centric traffic environments, and rapidly changing urban landscapes. Grab built GrabMaps to solve this problem, and adopted OpenAI's GPT-4o vision fine-tuning technology to further advance it.

Grab collected millions of street images through a driver network equipped with 360-degree cameras. It then fine-tuned GPT-4o by combining just 100 sample cases with map tiles, enabling the model to grasp context similarly to human workers even in situations involving complex road geometries or obscured signage.

Through this technology adoption, the following results were achieved:

• 20% improvement in lane count recognition accuracy
• 13% improvement in speed limit sign location accuracy (67% → 80%)
• Reduced operational costs and improved data reliability through decreased manual mapmaking work

Grab plans to continue expanding its AI capabilities going forward, including a multilingual voice assistant for the visually impaired and elderly, and an advanced support chatbot capable of handling complex inquiries.