AI research institute Ai2 has released new robotics models trained exclusively in simulations. The models are designed to work directly on real robots without any manually collected data or fine-tuning, what researchers call zero-shot sim-to-real transfer. The approach could significantly accelerate development: with conventional training, researchers typically needed months of teleoperated real-world demonstrations to make simulation-trained robots reliable.
The two new open-source systems are called MolmoSpaces and MolmoBot. MolmoSpaces includes over 230,000 indoor scenes, more than 130,000 curated objects, and over 42 million physics-based robotic grasping annotations. MolmoBot builds on this foundation and can pick up and place objects, open drawers, and operate doors, all without ever seeing real training data for these tasks.
According to Ranjay Krishna, director of the PRIOR team at Ai2, the gap between simulation and reality shrinks when researchers dramatically increase the variety of simulated environments, objects, and camera conditions. All models and tools are openly available, and technical details can be found in the paper.
