MolmoAct 2 shows what open models can unlock for robotics

Ten days. That's how long MolmoAct 2 had been out in the wild before it showed up at South Park Commons' embodied AI hackathon in San Francisco on May 15. By the end of the weekend, the winning build belonged to Binh Pham, a robotics software engineer at LiveKit, who hooked the model into a voice-controlled robot. His reaction when it worked, in his own words: "I really shouted, 'whoa is this actually working!'"

What Pham built wasn't just a demo that followed a script. He gave the robot jobs it had never specifically been trained to do, and it handled them anyway. The part that got him wasn't just that the model could spot objects in front of the robot—plenty of vision systems can do that. It was that MolmoAct 2 could take those observations and turn them into sensible action, connecting a spoken instruction to the actual scene and figuring out what to do about it.

That connective step is usually where robot learning falls apart. Recognizing an object is one thing; reasoning about where it sits in space and producing motor commands that make physical sense on real hardware is another problem entirely, and it's the one that's kept general-purpose robotics out of reach for a long time. Pham said the model's spatial awareness was strong enough that it identified objects reliably and then acted on that understanding without task-specific training.

His conclusion was blunt: this is the kind of generalization roboticists have chased for decades, and now it's sitting in the open, downloadable by anyone. "I truly think it's like the GPT-2 moment of robotics," he said. That comparison isn't about raw capability so much as timing—a signal that something foundational just became available to build on.

And MolmoAct 2 isn't shipped as a locked checkpoint you can only poke at from the outside. AI2 released the model weights, the training data, the training code, and fine-tuning scripts alongside it. In a field where reproducing someone else's robot learning results is often a nightmare, handing over the whole stack—not just the output—is what turns a single working demo into something a community can actually extend.