Allen Institute Unveils Simulation-Only Robot Training in Open-Source Push
Ai2's MolmoBot system trains robots entirely in virtual environments, outperforming physical-data models while releasing 230,000 scenes and 42 million annotations to the public.
The Allen Institute for AI has released a suite of open-source robotics tools that train machines to manipulate objects using only simulated environments, eliminating the need for expensive physical demonstrations and marking a potential shift in how the field approaches embodied AI development.
The institute's MolmoBot system and accompanying MolmoSpaces dataset enable robots to perform manipulation tasks with what researchers call zero-shot transfer—moving directly from virtual training to real hardware without additional fine-tuning. In benchmark tests, the approach outperformed models trained on physical data when deployed on Franka FR3 and Rainbow Robotics arms.
MolmoSpaces includes more than 230,000 indoor scenes, 130,000 object models, and 42 million grasp annotations distributed across multiple simulator platforms. The institute has made all components fully open-source, including underlying code and assets designed for reproducibility across different robot embodiments.
The release arrives as the artificial intelligence industry grapples with competing philosophies around openness. Meta has championed open-source models as accelerants for technological progress, while companies including OpenAI and Anthropic have argued that releasing foundational code creates unacceptable safety risks. Meta's own AI development has faced internal turbulence, with the company reportedly delaying its Avocado model and weighing whether to license Google's Gemini after disappointing trial runs.