AI2 Releases Robotics Models Trained Entirely in Simulation, Achieving Zero-Shot Real-World Transfer

AI research institute AI2 has released two open-source robotics models—MolmoSpaces and MolmoBot—trained exclusively in simulation that achieve zero-shot transfer to real robots without any manually collected real-world data or fine-tuning.

What's New

The models represent a significant shift in robotics training methodology. Conventional approaches require researchers to spend months collecting teleoperated real-world demonstrations before simulation-trained robots become reliable. AI2's approach eliminates this bottleneck entirely.

MolmoSpaces, the foundation dataset, contains:

• 230,000+ indoor scenes
• 130,000+ curated objects
• 42 million physics-based robotic grasping annotations

MolmoBot, built on MolmoSpaces, demonstrates capabilities including:

• Object picking and placement
• Opening drawers
• Operating doors

All tasks execute without training data from real-world demonstrations.

The Technical Approach

According to Ranjay Krishna, director of the PRIOR team at AI2, the key insight is straightforward: the simulation-to-reality gap shrinks dramatically when researchers increase the variety of simulated environments, objects, and camera conditions. Rather than improving physics simulation fidelity, the models benefit from diversity in training conditions.

This aligns with recent findings in robotics research showing that breadth of training distribution often matters more than pixel-perfect simulation accuracy. By exposing models to hundreds of thousands of variations in scene configuration, object types, and viewpoints, the models learn generalizable behaviors that transfer directly to physical systems.

Open-Source Release

Both models and supporting tools are available publicly. Technical details are available in the accompanying research paper. The open-source approach allows other research groups and robotics companies to build on the foundation rather than starting from zero with their own simulation data collection.

What This Means

This work addresses one of robotics' most significant friction points: the cost and time required to train deployable systems. If the zero-shot transfer results hold up in broader testing, the implications are substantial. Companies and research labs could dramatically reduce development timelines—from months of manual demonstration collection to weeks of model training. This could accelerate deployment of manipulation tasks in warehousing, manufacturing, and service robotics.

The emphasis on simulation diversity over physics accuracy also reframes how the robotics community should approach simulation tools. Rather than competing on fidelity, platforms that generate high-variance synthetic training data may prove more valuable. This could shift investment and resource allocation within the robotics software ecosystem.