Ai2: Building physical AI with virtual simulation data

Ai2's MolmoBot represents a significant advancement in physical AI development by leveraging virtual simulation data to train robotic manipulation systems, reducing reliance on costly and labor-intensive real-world demonstrations. Unlike traditional approaches that depend on extensive human-collected datasetssuch as Google DeepMinds RT-1, which required 130,000 episodes over 17 monthsMolmoBot is trained entirely on synthetic data, offering a more scalable and accessible model for the research community. This shift not only lowers research costs but also democratizes the development of generalist manipulation agents, enabling broader experimentation and innovation. The core innovation