The approach aims to make robotics development far more efficient. Under conventional methods, researchers often spend months collecting teleoperated real-world demonstrations to make simulation-trained robots perform reliably outside virtual environments.
The two new open-source systems are called MolmoSpaces and MolmoBot.
MolmoSpaces includes more than 230,000 indoor scenes, over 130,000 curated objects, and more than 42 million physics-based robotic grasp annotations.
Built on that foundation, MolmoBot can grasp and place objects, as well as operate drawers and doors, despite never having seen real-world training data for those tasks.
According to Ranjay Krishna, director of Ai2’s PRIOR team, the gap between simulation and reality can shrink significantly when the diversity of simulated environments, objects, and camera conditions is scaled up aggressively.
Ai2 has made the full set of models and tools openly available, with additional technical details provided in the accompanying research paper.
Conclusion
Ai2’s release suggests that robotics may be moving toward a more scalable training model, where simulation does far more of the heavy lifting. If zero-shot sim-to-real transfer proves reliable at scale, it could sharply reduce the cost and time required to build useful robotic systems.
ES 
EN