The Allen Institute for AI (Ai2) has released MolmoBot, an open robotic manipulation model suite trained entirely on synthetic simulation data โ€” no real-world demonstrations required.

Zero-Shot, No Fine-Tuning

The core claim: MolmoBot transfers directly from simulation to real robots without any additional real-world data or fine-tuning. This "zero-shot sim-to-real" milestone has been a long-standing goal in robotics research. Most current systems โ€” including Google DeepMind's RT-1 (130,000 teleoperated episodes) and Physical Intelligence's ฯ€0 series โ€” still depend on large volumes of expensive, manually collected robot data.

Ai2's bet is the opposite. Rather than patching the sim-to-real gap with real-world data, they dramatically expanded simulation diversity: 1.8 million procedurally generated manipulation trajectories across varied objects, lighting, viewpoints, and physics parameters.

What's Included

MolmoBot runs on two robot platforms โ€” the Rainbow Robotics RB-Y1 mobile manipulator and the Franka FR3 tabletop arm. It handles pick-and-place tasks, drawer/cabinet opening, and door manipulation on unseen objects and environments. Performance is reported as competitive with ฯ€0 and ฯ€0.5.

Underpinning it all is MolmoSpaces, an open simulation ecosystem with 230,000+ indoor scenes, 130,000+ object models, and 42 million physics-grounded grasp annotations.

Fully Open

Everything โ€” training data, data generation pipelines, training code, and a technical report โ€” is released openly. Ai2's goal is to lower the barrier for academic labs that lack large-scale teleoperation setups.

"Our latest advancement shifts the constraint in robotics from collecting manual demonstrations to designing better virtual worlds," said Ranjay Krishna, Director of Ai2's PRIOR team. "And that's a problem we can solve."