Ai2 Rethinks Robotics Training With a Simulation-First Approach to Physical AI

Researchers at the Allen Institute for AI (Ai2) say they’ve achieved a breakthrough in physical AI: zero-shot sim-to-real transfer. This milestone allows robotic models trained entirely in simulation to perform tasks in the real world without requiring any real-world data, making it significantly easier to move knowledge from virtual training to actual robots.

To mark this milestone, the nonprofit lab is unveiling MolmoSpaces and MolmoBot—two open tools designed to help researchers, roboticists, and anyone exploring physical AI harness and build on its technology.

“Most approaches try to close the sim-to-real gap by adding more real-world data,” Ranjay Krishna, Ai2’s director of its PRIOR team, is quoted as saying in a blog post. “We took the opposite bet: that the gap shrinks when you dramatically expand the diversity of simulated environments, objects, and camera conditions. Our latest advancement shifts the constraint in robotics from collecting manual demonstrations to designing better virtual worlds, and that’s a problem we can solve.”

Traditionally, physical AI development relied on training models in simulation and then retraining them with real-world data—a slow and laborious process. Ai2’s approach removes that bottleneck, potentially accelerating robotics research and shortening time to real-world deployment. While other organizations, including DeepMind, OpenAI, Nvidia, and Meta, are exploring zero-shot sim-to-real techniques, Ai2 stands out for making its work fully open-source.

Meet MolmoSpaces and MolmoBot

With MolmoSpaces and MolmoBot, Ai2 has turned its research into practical, openly accessible tools that researchers and developers can integrate directly into their own projects.

The former is an open ecosystem for embodied AI research featuring more than 230,000 indoor scenes, more than 130,000 curated object assets, and over 42 million physics-grounded robotic grasp annotations. These were all curated from Objaverse and THOR, the organization’s environment that generates data from synthetic worlds for training robotic policies. Developers can leverage MolmoSpaces’ assets in other common simulators, including MuJoCo, ManiSkill, and Nvidia’s Isaac Lab/Sim, using a USD conversion script.

Along with the aforementioned components, MolmoSpaces also features MolmoSpaces-Bench, a benchmark for evaluating generalist policies with a focus on “generalization under systematic, controlled variation.” In other words, can the robot act intelligently in new, slightly different situations it hasn’t encountered before?

MolmoBot, on the other hand, is an open manipulation model suite trained entirely on synthetic data. Built on MolmoSpaces, this model serves as Ai2’s proof in the pudding, showing that simulation-trained robots can perform real-world tasks without any additional real-world data. Tested on the Rainbow Robotics RB-Y1 mobile manipulator and the Franka FR3 tabletop arm, MolmoBot successfully handled everyday tasks such as picking up objects, opening drawers and cabinets, and navigating doors, all with items and environments it had never encountered before. This suggests the model can adapt and perform in real-world situations, showing a step toward practical, flexible robotics that could eventually assist in homes, workplaces, and research labs.

Powering MolmoBot is MolmoBot-Data, a massive dataset built from millions of simulated robot movements. By training across a huge variety of objects, environments, angles, lighting, and conditions, the system learns to handle tasks in ways that can transfer directly to the real world. And although there is one dataset, it also has three different policy architectures, “blueprints” of a robot’s decision-making system:

MolmoBot: the primary vision-language model manipulation policy built on Molmo2. Ai2 claims it’s the highest performer across its evaluations.

MolmoBot-SPOC: A lightweight transformer policy that’s adapted from the original Shortest Path Oracle Clone (SPOC) navigation architecture. It provides competitive performance with “significantly” fewer parameters. Ai2 describes this best for compute-constrained settings.

MolmoBot-Pi0: This final policy architecture uses the same model design as Physical Intelligence’s π0 system, allowing researchers to evaluate robots trained on simulation versus those trained on real-world data.

“We see MolmoBot as a test of whether fully synthetic training can work for manipulation,” Ai2 writes in a blog post. “Our results suggest it can, without expensive real-world data collection, task-specific fine-tuning, photorealistic rendering, or complex domain adaptation. The practical outcome is that the bottleneck moves from manually collecting data to designing better virtual worlds—a problem we can scale with compute and open infrastructure.”

The Quest for the ‘Robot Brain’ Continues

MolmoSpaces and MolmoBot are the latest expansion to Ai2’s line of open-source vision-language models, especially in the field of robotics. In August, the nonprofit lab debuted MolmoAct, which trains robots to move through the world with greater spatial awareness. It followed that up with the release of Molmo2 in December, a more powerful model that supports multiple images and video, as well as grounding.

This week’s announcement seems to confirm a prediction Ai2 researcher Jiafei Duan shared with me last summer: that 2026 would be the year when physical, or embodied AI, takes center stage. By achieving zero-shot sim-to-real transfer, Ai2 has eliminated a major bottleneck in robotics research, showing that robots can now learn entirely in simulation and succeed in the real world. Coming days ahead of this year’s Nvidia GTC event, the news could accelerate the development of practical robots capable of performing complex tasks outside the lab, ushering in a shift in how the field approaches training intelligent machines.

“Our mission is to build AI that advances science and expands what humanity can discover,” Ali Farhadi, Ai2’s chief executive, remarked in a statement. “Robotics can become a foundational scientific instrument, helping researchers move faster and explore new questions. To get there, we need systems that generalize in the real world and tools the global research community can build on together.”

Ai2 has built a hands-on playground for developers to explore MolmoSpaces. Its dataset is also available on Hugging Face, while its code is open-sourced on GitHub.

Featured Image: AI-generated image of a virtual robotics training environment showing a robot navigating and manipulating objects inside a simulated home interior. Credit: Adobe Firefly