Breakthrough in Sim-to-Real Transfer

Ai2 has demonstrated a significant milestone in robotics: achieving zero-shot transfer from simulation to real robots without the months of teleoperated real-world demonstrations previously required. This challenges a fundamental assumption in the field—that collecting proprietary real-world datasets at scale was necessary for reliable robot behavior in practice.

The breakthrough centers on a counterintuitive insight: instead of adding more real-world data to close the sim-to-real gap, Ai2 dramatically expanded the diversity of simulated environments, objects, lighting conditions, and task definitions. The research shows that breadth in simulation leads to stronger generalization to real-world scenarios.

MolmoSpaces: Open Simulation Infrastructure

MolmoSpaces is a large-scale, open simulation ecosystem designed for embodied AI research. It includes:

• 230,000+ indoor scenes for training and evaluation
• 130,000+ curated object assets with varied properties and articulations
• 42+ million physics-grounded robotic grasp annotations
• Support for systematic variation of object properties, layouts, lighting, and task definitions
• Compatibility with widely-used simulators: MuJoCo, NVIDIA Isaac Lab, and NVIDIA Isaac Sim

By releasing these assets and tools openly, Ai2 transforms simulation from a proprietary pretraining tool into shared scientific infrastructure, making robotics research more reproducible and accessible.

MolmoBot: Open Manipulation Model Suite

MolmoBot is built on MolmoSpaces and demonstrates the practical impact of the zero-shot approach. Trained entirely on synthetic data, it performs:

• Pick-and-place operations on unseen objects
• Articulated object manipulation (opening drawers, cabinets)
• Door opening tasks in new environments
• Deployment across multiple robot systems, including mobile manipulators

Notably, MolmoBot achieves this without photorealistic rendering, real-world demonstrations, or task-specific fine-tuning—demonstrating that diversity matters more than scale.

Implications for Robotics Research

This shift has substantial implications. By prioritizing simulation diversity over manual data collection, the bottleneck in robotics research moves from gathering months of teleoperated demonstrations to designing richer virtual environments. This makes physical AI more accessible to smaller labs and organizations without large-scale real-world data collection capabilities.

All components—models, simulation infrastructure, grasp annotations, data generation pipelines, and benchmarking tools—are released as open source, enabling the broader research community to build on, test, and improve the work.