Zero-Shot Sim-to-Real Transfer

Ai2 has announced a significant breakthrough in robotics: models trained entirely in simulation that successfully transfer to real-world robots with zero-shot transfer—meaning no additional manually-collected real-world data or task-specific fine-tuning required. This challenges a core assumption in the robotics field, which has historically required months of teleoperated demonstrations to achieve reliable real-world performance after simulation training.

MolmoSpaces: Large-Scale Simulation Infrastructure

MolmoSpaces is an open ecosystem for embodied AI research providing researchers with unprecedented simulation diversity:

• 230,000+ indoor scenes with varied layouts and configurations
• 130,000+ curated object assets for comprehensive coverage
• 42+ million physics-grounded robotic grasp annotations to ground learning in realistic physics
• Support for controlled variation of object properties, lighting, articulation, physics parameters, and task definitions
• Compatibility with popular simulators including MuJoCo, NVIDIA Isaac Lab, and NVIDIA Isaac Sim

The key insight: sufficient diversity in simulation alone can produce robust real-world behavior, eliminating the traditional bottleneck of expensive real-world data collection.

MolmoBot: Practical Manipulation Capabilities

Built on MolmoSpaces, MolmoBot is a fully open manipulation model suite trained on synthetic data only. It demonstrates zero-shot transfer across:

• Two different robot systems (including a mobile manipulator)
• Pick and place tasks on unseen objects
• Articulated object manipulation (opening drawers, cabinets, doors)
• Novel environments without task-specific adaptation

Notably, the system achieves this without photorealistic rendering, which is a common approach in other sim-to-real research.

Paradigm Shift in Robotics Research

This work reframes the robotics research constraint: instead of investing in manual data collection, researchers can focus on designing richer simulated environments that scale with compute. The research shows that breadth in simulation leads to stronger generalization in reality—diversity matters more than scale alone.

The fully open release includes models, simulation infrastructure, grasp annotations, data generation pipelines, and benchmarking tools, designed to integrate into existing research ecosystems and accelerate progress across the community.