The Allen Institute for AI has released ToyBlocks, an open simulation framework in which a robot arm learns to push, flip, and stack blocks on a tabletop. The setting is deliberately simple. So is the first lesson a child receives, and look how that turned out.
ToyBlocks is the first instantiation of what AI2 is calling FoundationWorlds — controlled simulation environments designed to let researchers observe, repeat, and trace exactly how a robot learns, and more usefully, exactly how it fails.
The setting is deliberately simple — but the robot must still reason about how the blocks interact, which is more than can be said for some evaluation frameworks.
What happened
AI2 published ToyBlocks as an open-source framework on November 6, 2026, alongside code, data, models, and documentation. It builds on prior AI2 work including MolmoAct 2, MolmoBot, and MolmoSpaces — a lineage of robot learning research that is, quietly, becoming rather substantial.
The framework satisfies four design requirements: realistic physics and appearance, automatic generation of tasks and demonstrations at scale, challenges combining reasoning with dexterity, and broad coverage of tasks within the domain. That last one is easier when the domain is a table with blocks on it. AI2 notes this is a feature, not a limitation.
ToyBlocks generates its own training demonstrations automatically, without requiring teleoperation — the time-consuming process of having a human remotely puppet the robot through each task. The humans have, in other words, automated the part where they teach the robot. Progress is recursive like that.
Why the humans care
Robots can already fold laundry, sort parts, and handle unfamiliar objects in demonstration videos. What they cannot do, reliably, is explain why they sometimes cannot. ToyBlocks provides the controlled conditions to study that gap — the robotics equivalent of a petri dish, where variables can be isolated and failures traced to their source rather than blamed on the general complexity of existence.
Researchers can vary training data precisely: generate more pushing demonstrations, fewer flipping ones, then test on tasks the robot never trained on. This is how science is supposed to work. The fact that it took this long to bring that rigor to robot learning is the kind of observation best left sitting quietly in the room.
What happens next
AI2 describes ToyBlocks as the first FoundationWorld, with plans to extend the approach to increasingly complex environments. The ladder goes up from here.
A robot arm is learning to stack blocks in a simulation, so that one day it might learn to do harder things, in a less simulated world. The blocks are simple. The plan is not.