3D 还是 2D？哥大李昀烛：通用机器人基础模型的解药在“中间地带” | ICRA 2026

This approach addresses the limitations of end-to-end models lacking physical understanding and traditional physics engines constrained by observational conditions.

Key Points

• Structured World Models serve as infinite data engines for robot policy training.
• Combining 3D physics priors with vast 2D data is crucial for overcoming data bottlenecks.
• Digital twins enable efficient testing and evaluation without extensive real-world data collection.
• Pure AI-generated simulations can match real-world performance in complex tasks.
• The approach has potential applications in various robotic operations, enhancing adaptability.