3D or 2D? Columbia's Li Yunzhu: The antidote to general-purpose robot foundation models lies in the 'middle ground' | ICRA 2026
Columbia University Assistant Professor Yunzhu Li presented a novel approach at ICRA 2026, proposing "Structured World Models" as a scalable data engine for robot policy training and evaluation. This method aims to bridge the gap between purely data-driven end-to-end models and physics-based simulators by integrating 3D physical priors with extensive 2D data learning. The proposed digital twin framework allows for efficient, high-fidelity simulation of robot-environment interactions, significantly accelerating the testing and refinement of AI policies compared to real-world robot trials. AI
IMPACT Accelerates robot policy development by enabling efficient, high-fidelity simulation, reducing reliance on costly real-world testing.