Ai2 Revolutionizes Physical AI with MolmoBot and Simulated Data

Ai2 and the Rise of Physical AI through Virtual Simulation

The Allen Institute for AI (Ai2) is redefining the development of physical AI by leveraging virtual simulation data, with initiatives like MolmoBot. Traditionally, training robotic systems required costly and manual demonstrations. Technology providers developing general-purpose manipulation agents often consider extensive real-world training essential for these systems.

To illustrate this approach, projects like DROID have gathered 76,000 teleoperated trajectories from 13 institutions, representing about 350 hours of human effort. Meanwhile, Google DeepMind's RT-1 project required 130,000 episodes collected over 17 months by human operators. This reliance on manual and proprietary data collection inflates research budgets and concentrates capabilities within a small group of well-funded industrial laboratories.

"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 fundamental scientific tool, helping researchers move faster and explore new questions. To achieve this, we need systems that generalize in the real world and tools that the global research community can collaborate on. Demonstrating the transfer from simulation to reality is a significant step in this direction."

MolmoBot-Data: A Breakthrough in Data Collection

Researchers at the Allen Institute for AI (Ai2) are proposing a different economic model with MolmoBot, a suite of robotic manipulation models fully trained on synthetic information. By procedurally generating trajectories within a system called MolmoSpaces, the team avoids the need for human teleoperation.

The associated dataset, MolmoBot-Data, contains 1.8 million expert manipulation trajectories. This collection was produced by combining the physics engine MuJoCo with aggressive domain randomization, varying objects, viewpoints, lighting, and dynamics.

"Most approaches try to close the gap between simulation and reality by adding more real-world data," said Ranjay Krishna, director of the PRIOR team at Ai2. "We took the opposite bet: that the gap narrows when you significantly broaden the diversity of simulated environments, objects, and camera conditions. Our latest advancement shifts the constraint in robotics from collecting manual demonstrations to designing better virtual worlds, and that's a problem we can solve."

Generating Virtual Simulation Data for Physical AI

Using 100 Nvidia A100 GPUs, the pipeline created approximately 1,024 episodes per hour of GPU time, equivalent to over 130 hours of robotic experience for every hour of real-time. Compared to real-world data collection, this represents nearly four times the data throughput, directly impacting the project's return on investment by accelerating deployment cycles.

The MolmoBot suite includes three distinct classes of policies evaluated on two platforms: the mobile manipulator Rainbow Robotics RB-Y1 and the tabletop arm Franka FR3. The main model, built on a vision-language foundation called Molmo2, processes multiple steps of RGB observations and language instructions to dictate actions.

Hardware Flexibility with Ai2's MolmoBot

For edge computing environments where resources are limited, researchers propose MolmoBot-SPOC, a lightweight transformer policy with fewer parameters. MolmoBot-Pi0 uses a PaliGemma base to match the architecture of the π0 model from Physical Intelligence, allowing for direct performance comparisons.

In physical testing, these policies demonstrated zero-shot transfer to real-world tasks involving unseen objects and environments without any fine-tuning.

In tabletop grasping and placement evaluations, the main MolmoBot model achieved a success rate of 79.2%. This surpasses the π0.5 model, which was trained on real-world demonstration data, achieving a success rate of 39.2%. For mobile manipulation, the policies successfully executed tasks such as approaching, grasping, and pulling doors across their full range of motion.

Providing these varied architectures allows organizations to integrate high-performing physical AI systems without being locked into a single proprietary vendor ecosystem or extensive data collection infrastructure.

The open publication of the entire MolmoBot stack—including training data, generation pipelines, and model architectures—enables internal auditing and adaptation. Anyone exploring physical AI can leverage these open tools for simulation and building capable systems while controlling costs.

"For AI to truly advance science, progress cannot depend on closed data or isolated systems," continues Ali Farhadi, CEO of Ai2. "It requires a shared infrastructure on which researchers worldwide can build, test, and improve together. This is how we believe physical AI will advance."