Ai2's MolmoBot Advances Physical AI with Synthetic Simulation Data

Virtual simulation data is emerging as a powerful driver in the development of physical artificial intelligence (AI) systems, according to new research by the Allen Institute for AI (Ai2). The institute’s latest project, MolmoBot, introduces an open-source robotic manipulation suite trained exclusively on synthetically generated data, marking a shift from traditional reliance on real-world demonstrations.

Historically, training robots to interact with physical environments has depended on expensive, labor-intensive data collection, with models developed using proprietary datasets from resource-rich industrial labs. For example, Google DeepMind’s RT-1 model required 130,000 episodes over 17 months, while the DROID project compiled 76,000 teleoperated trajectories across 13 institutions. This approach has driven up costs and limited access to advanced robotics capabilities.

Aiming to democratize access and reduce costs, Ai2 developed MolmoBot using procedural generation tools and physics-based simulations. The project leverages MolmoSpaces, Ai2’s custom system for generating manipulation tasks, to create large datasets procedurally, eliminating the need for time-consuming human teleoperation.

The resulting MolmoBot-Data set includes 1.8 million expert manipulation trajectories, created by combining the MuJoCo physics engine with extensive domain randomisation—altering objects, lighting, and camera perspectives to better expose the model to environmental diversity. This strategy helps improve the so-called ‘sim-to-real’ transfer, where models trained in virtual worlds are deployed effectively on physical hardware.

"Our mission is to build AI that advances science and expands what humanity can discover," said Ali Farhadi, CEO of Ai2. "Robotics can become a foundational scientific instrument. To get there, we need systems that generalise in the real world and tools the global research community can build on together. Demonstrating transfer from simulation to reality is a meaningful step in that direction."

To generate such a large quantity of training data, the team employed 100 Nvidia A100 GPUs, achieving over 1,000 episodes per GPU-hour—yielding over 130 hours of simulated robot experience for every hour of actual time. This represents a nearly fourfold increase in data throughput compared to real-world collection, enabling much faster iteration and reduced operational costs.

MolmoBot comprises three policy classes, each tested across two hardware platforms: the Rainbow Robotics RB-Y1 mobile manipulator and the Franka FR3 tabletop arm. The main model incorporates a Molmo2 vision-language backbone, allowing it to process sequences of RGB images and natural language instructions to control actions.

For edge environments with limited resources, Ai2 offers MolmoBot-SPOC, a lighter-weight transformer variant, and MolmoBot-Pi0, which uses a PaliGemma backbone for direct comparison with Physical Intelligence's π0 model. These policies, evaluated on real-world tasks involving unfamiliar objects and settings, demonstrated strong ‘zero-shot’ performance—transferring from simulation to reality without additional fine-tuning.

In tabletop pick-and-place evaluations, MolmoBot achieved a 79.2% success rate, outperforming real-world-trained models such as π0.5 (which managed 39.2%). The models also successfully performed mobile manipulation tasks such as approaching and opening doors.

According to Ranjay Krishna, Director of the PRIOR team at Ai2, "Most approaches try to close the sim-to-real gap by adding more real-world data. We took the opposite bet: that the gap shrinks when you dramatically expand the diversity of simulated environments."

By releasing the full MolmoBot stack—including training data, pipelines, and model architectures—as open-source, Ai2 aims to facilitate broader adoption and independent auditing of physical AI models. Organisations can experiment, adapt, and scale robotics solutions without becoming reliant on closed vendor ecosystems or prohibitively expensive data processes.

"For AI to truly advance science, progress cannot depend on closed data or isolated systems," said Farhadi, highlighting the importance of open and collaborative infrastructure for future advancements in physical AI.

artificialintelligence-news.com