Virtual simulation data is driving the development of physical AI across enterprise environments through initiatives such as Ai2’s MolmoBot.
Telling hardware to interact with the real world has traditionally relied on very expensive, manually assembled demonstrations. Technology providers who build generalist operational agents typically assemble extensive real-world training as the basis for these systems.
For context, a project like DROID includes 76,000 teleoperated trajectories collected across 13 institutions, which equates to approximately 350 hours of human effort. Google DeepMind’s RT-1 required 130,000 episodes collected over 17 months by human operators. This reliance on unique manual data collection inflates research budgets and concentrates capacity within a small group of resource-rich industrial laboratories.
“Our mission is to build AI that advances science and expands what humanity can discover,” said Ali Farhadi, CEO of Ai2. “Robotics will become a fundamental scientific instrument, helping researchers move faster and explore new questions. To get there, we need systems that can be generalized in the real world and tools that the global research community can build together. Demonstrating the transition from simulation to reality is a meaningful step in that direction.”
Researchers at the Allen Institute for AI (Ai2) offer a different economic model using MolmoBot, a suite of open robot interaction models trained entirely on synthetic information. By procedurally generating trajectories within a system called MolmoSpaces, the team avoids the need for remote human control.
The accompanying dataset, MolmoBot-Data, contains trajectories of 1.8 million expert actions. This collection was created by combining the MuJoCo physics engine with aggressive domain randomization and a variety of objects, perspectives, lighting, and dynamics.
“Most approaches try to bridge the gap between simulation and reality by adding more real-world data,” said Ranjay Krishna, director of Ai2’s PRIOR team. “We made the opposite bet: the gap would narrow if we dramatically expanded the diversity of simulated environments, objects, and camera conditions. Our latest advances shift the constraints of robotics from collecting manual demonstrations to better designing virtual worlds. This is a problem we can solve.”
Generation of virtual simulation data for physical AI
Using 100 Nvidia A100 GPUs, the pipeline produced approximately 1,024 episodes per GPU hour. This equates to more than 130 hours of robot experience in every measured hour.
This equates to nearly four times the data throughput compared to actual data collection, directly impacting a project’s return on investment by accelerating deployment cycles.
The MolmoBot suite includes three different policy classes that were evaluated on two platforms: the Rainbow Robotics RB-Y1 mobile manipulator and the Franka FR3 tabletop arm. The main model, built on the Molmo2 vision language backbone, processes multiple timesteps of RGB observations and language instructions to direct actions.
Hardware flexibility with Ai2’s MolmoBot
For edge computing environments with limited resources, researchers offer MolmoBot-SPOC, a lightweight transformer policy with fewer parameters. MolmoBot-Pi0 uses the PaliGemma backbone to match the architecture of Physical Intelligence’s π0 model, allowing direct performance comparisons.
During physical testing, we demonstrated that these policies can be zero-shot transferred to real-world tasks involving invisible objects and environments without any fine-tuning.
In a tabletop pick-and-place evaluation, the leading MolmoBot model achieved a success rate of 79.2%. This beats π0.5, a model trained on extensive real-world demonstration data, which achieved a 39.2% success rate. For mobile operations, the policy successfully performed tasks such as approaching, grabbing, and pulling the door throughout its operating range.
These diverse architectures allow organizations to integrate capable physical AI systems without being tied to a single proprietary vendor ecosystem or extensive data collection infrastructure.
Open release of the entire MolmoBot stack, including training data, generation pipeline, and model architecture, enables internal auditing and adaptation. Anyone researching physical AI can leverage these open tools to simulate and build capable systems while controlling costs.
“For AI to truly advance science, that progress cannot depend on closed data or isolated systems,” continued Ali Farhadi, CEO of Ai2. “That requires a shared infrastructure that allows researchers around the world to build, test, and improve together. We believe this is the way physical AI will move forward.”
SEE ALSO: New partnership brings smart robots to hazardous environments
Want to learn more about AI and big data from industry leaders? Check out the AI & Big Data Expos in Amsterdam, California, and London. This comprehensive event is part of TechEx and co-located with other major technology events such as Cyber Security & Cloud Expo. Click here for more information.
AI News is brought to you by TechForge Media. Learn about other upcoming enterprise technology events and webinars.
