Blueprint-Bench 2 Released: AI's First Real-World Spatial Reasoning Based on 2D Photos

CryptoWorld News: Andon Labs has released Blueprint-Bench 2, an evaluation benchmark designed to test AI spatial reasoning capabilities, requiring agents to generate accurate 2D floor plans from approximately 20 indoor photos. This test introduces a cross-task “notebook” system that allows agents to record layout rules and optimize their generation strategies while processing 50 apartment units. The evaluation results show that gpt-5.5, gemini 3.1 pro, and claude opus 4.7 rank in the top three, marking the first time they have demonstrated the ability to understand real spaces from 2D photos. For example, gemini 3.1 pro can infer the camera’s shooting orientation from washing machines in different photos, while gpt-5.5 infers the connectivity between bedrooms and hallways from the positions of door frames. Meanwhile, the performance of gemini robotics-er 1.6 fell short of expectations, with a score even lower than the lightweight gemini 3 flash, indicating that its spatial strengths have not been successfully translated into floor plan generation capabilities.