The Allen Institute for AI (Ai2) has released MolmoBot, an open robotic manipulation model trained entirely on synthetic simulation data that transfers directly to real-world robots with zero fine-tuning. The breakthrough challenges a long-held assumption in robotics that virtual training alone is not sufficient for real-world performance.

MolmoBot was trained on 1.8 million simulated robot trajectories spanning over 100,000 environments and 30,000 unique objects. In testing, the model successfully operated a Franka FR3 robotic arm and a Rainbow Robotics RB-Y1 mobile manipulator, performing manipulation tasks on unseen objects without any real-world data collection.

Ai2 has open-sourced the entire stack, including MolmoSpaces, a large-scale simulation ecosystem for embodied AI. The release could dramatically reduce the cost and time required to train capable robots, eliminating the expensive real-world data collection bottleneck that has slowed the robotics industry.