The Allen Institute for AI has released MolmoSpaces – an open platform for what the industry calls embodied AI, or «embodied artificial intelligence». Simply put, this is AI that doesn't just generate text or images but is capable of controlling physical devices: robots, manipulators, and drones.
What Is Embodied AI and How It Differs from Standard AI Models
What Is Embodied AI and Why Is It a Separate Category
Ordinary language models work in the digital space. They process text, answer questions, and write code. But if you need a model to control a robot – say, pick up a cup from a table or cross a room – a completely different skillset is required.
A robot must understand images from a camera, estimate distance to objects, plan movements, and adjust them in real time. This isn't just a question of increasing the number of model parameters, but of architecture, data, and training methods.
Previously, highly specialized systems were used for such tasks: one model recognized objects, another planned the trajectory, and a third managed motor skills. MolmoSpaces offers a different approach: using a multimodal model that is simultaneously capable of both seeing and acting.
Key Features and Components of the MolmoSpaces Platform
What Is Included in MolmoSpaces
The platform includes several components. First, there is the Molmo model itself: it already knows how to work with images and text, and now it has been adapted for controlling robots.
Second is the training dataset. For a model to learn to act in the physical world, it needs examples: video from robot cameras, recordings of movement trajectories, and action annotations. The Allen Institute has collected such data and made it publicly available.
Third is the testing infrastructure. Developers can check their models in simulators and then transfer them to real robots. This lowers the barrier to entry: you don't need to buy expensive equipment right away to start experiments.
The Role of Open Source Development in Embodied Artificial Intelligence
Why Openness Matters in This Field
Embodied AI is an expensive field. It requires robots, sensors, and significant computing power. Most research is conducted by large companies, and their results are rarely published in full. This slows down progress: every team is forced to solve the same basic problems from scratch.
MolmoSpaces is betting on a different development model. All components – the model, data, and code – are available for use and modification. This allows researchers and startups to experiment without starting from square one.
For the industry, this could mean accelerated progress. If more teams can work on embodied AI, more solutions will appear for warehouse logistics, home automation, and medical robotics. So far, these areas are developing slowly precisely because of high barriers to entry.
Future Outlook for MolmoSpaces and Robot Learning Systems
What's Next
The project has just launched, and it is too early to judge how effectively the model handles complex tasks. Embodied AI is not just technology but also engineering: even a high-quality model can glitch if the robot is poorly calibrated or if the data was collected in conditions that differ from reality.
But the very fact that an open platform has appeared is already an important step. Previously, most tools for embodied AI were either closed or too narrowly specialized. MolmoSpaces aims to create an ecosystem where different teams can work on a common task.
If this approach pays off, in a few years we will see robots that truly understand the surrounding world not through rigid algorithms, but thanks to learning from examples. For now, this is more of a research base than a ready-made solution, but that is exactly how serious technological changes usually begin.