Ai2 generated 1.8 million robotic trajectories without using real-world demonstrations. The scale highlights a shift in how physical AI systems are trained, reducing reliance on costly human-operated data collection.
The Allen Institute for AI developed MolmoBot, a robotic manipulation model suite trained entirely on synthetic data. Using its MolmoSpaces simulation system, the team generated diverse interaction scenarios across objects, lighting, and environments. The pipeline ran on 100 Nvidia A100 GPUs, producing over 130 hours of robot experience per hour of compute time.
Can Simulation Data Replace Real-World Training In Robotics?
Traditional approaches rely heavily on manual datasets, which remain resource-intensive and limited in scale. Google DeepMind’s RT-1 required 130,000 real-world episodes collected over 17 months, while the DROID project logged 76,000 trajectories across 13 institutions. In contrast, Ai2’s synthetic pipeline achieved nearly four times the data throughput.
Ai2’s models demonstrated zero-shot transfer, performing tasks in real-world environments without additional training. In tabletop pick-and-place tests, the primary MolmoBot system achieved a 79.2% success rate, compared with 39.2% for π0.5, a model trained on extensive physical demonstrations. The suite also executed mobile manipulation tasks such as door handling across full motion ranges.
“Our mission is to build AI that advances science and expands what humanity can discover,” said Ali Farhadi, CEO of Ai2.
He added that demonstrating transfer from simulation to real-world environments marks a key step toward scalable and collaborative robotics research.
Still, the open release of MolmoBot’s data and models introduces a different competitive dynamic, allowing broader access beyond well-funded labs. The next catalyst will be whether enterprises adopt simulation-first pipelines as a standard for deploying physical AI systems at scale.
