Robotics development has long depended on humans physically demonstrating tasks to machines. Ai2’s MolmoBot project proposes a different route: train manipulation policies entirely in simulation, then transfer them directly to real hardware without fine-tuning.
The economic difference is substantial. The DROID project gathered 76,000 teleoperated trajectories across 13 institutions, representing about 350 hours of human effort. Google DeepMind’s RT-1 used 130,000 episodes collected by operators over 17 months. Such programmes concentrate advanced robotics inside laboratories that can afford hardware, staff and long collection cycles.
Ai2 generated 1.8 million expert trajectories procedurally in MolmoSpaces. By releasing the data, generation pipeline and models, it wants to make physical AI research accessible beyond a small group of industrial labs.
More varied worlds can narrow the reality gap
Simulation is cheaper and faster, but simulated objects and physics never match reality perfectly. Many robotics programmes address that “sim-to-real” gap by adding real-world demonstrations. Ai2 made the opposite bet: expand the range of virtual conditions until the policy learns what variations do not matter.
MolmoSpaces combines the MuJoCo physics engine with aggressive domain randomisation. Objects, viewpoints, lighting and dynamics change across generated episodes. Instead of memorising one clean virtual scene, the system must learn behaviour robust enough to survive unfamiliar conditions.
Using 100 Nvidia A100 GPUs, the pipeline produced about 1,024 episodes per GPU-hour. Ai2 equates that to more than 130 hours of robot experience for every hour of elapsed time and nearly four times the throughput of real-world collection.
This moves the bottleneck. Teams need less time operating robots, but much more care designing virtual environments that represent the diversity and failure modes a machine will encounter. Poor simulation variety can create a policy that looks capable until a camera angle, surface or object behaves differently.
MolmoBot transfers directly to physical machines
The suite includes three policy classes tested on a Rainbow Robotics RB-Y1 mobile manipulator and a Franka FR3 tabletop arm. The main model uses a Molmo2 vision-language backbone to interpret multiple RGB observations and language instructions before choosing actions.
MolmoBot-SPOC offers a smaller transformer policy for constrained edge hardware. MolmoBot-Pi0 uses a PaliGemma backbone aligned with Physical Intelligence’s π0 architecture, enabling a more direct comparison between approaches.
In physical tests, the policies handled unseen objects and environments without additional training. The main MolmoBot achieved 79.2% success in tabletop pick-and-place tasks, compared with 39.2% for π0.5, which had been trained with extensive real-world demonstrations. Mobile tests included approaching, grasping and pulling doors through their full range.
The result does not prove simulation will replace physical testing. Robots still need evaluation around people, unusual materials and imperfect hardware. It does show that synthetic training can provide much more than a preliminary model waiting for a large real-world dataset.
Open simulation changes who can build physical AI
Ai2’s release includes architectures for different computing constraints rather than one closed package. Researchers and companies can audit the training process, adapt environments and compare policies without accepting a single vendor’s infrastructure.
That openness is strategically important. Synthetic data can reduce labour and hardware costs, but a proprietary generation system could merely replace one barrier with another. Shared tools let more teams reproduce findings and create simulations for specialised scientific or industrial tasks.
The next challenge is proving breadth. A strong result on pick-and-place and door manipulation must extend to deformable objects, clutter, tool use and safety-critical interaction. Organisations should also measure where synthetic training fails, not only average success.
MolmoBot nevertheless reframes the data question in robotics. Progress may depend less on collecting an ever-larger archive of demonstrations and more on designing virtual worlds that teach robust concepts. That makes simulation an active research instrument rather than a cheaper imitation of reality.
If the approach continues to transfer across hardware and tasks, physical AI development could become faster and less concentrated. More laboratories would be able to train capable systems, test hypotheses and contribute improvements. Ai2’s most consequential result may therefore be not one robot’s performance, but a credible open path toward producing the experience that many future robots will need.
