Groundbreaking ESI-Bench Reveals AI's 'Action Blindness' in 3D Environments

A Stanford-led team unveiled ESI-Bench, highlighting state-of-the-art AI's struggle with 'action blindness' in 3D tasks. Despite advanced capabilities, AI models falter in selecting informative actions needed for spatial reasoning. The benchmark, built on existing 3D simulations, establishes a new standard for evaluating multimodal AI. This shift signals crucial implications for robotics and physical-world applications.