Cover: AI-generated editorial illustration.
On November 6, 2026, the Allen Institute for AI published ToyBlocks, the first release under its FoundationWorlds initiative. The idea is simple in form but practically important: give robotics researchers a fully controlled simulation where they can vary one thing at a time, run the test again, and trace failures back to a cause.
The setup is a robot arm and a set of lettered blocks on a tabletop. That narrow scope is intentional. Allen AI argues that broad manipulation benchmarks make it hard to tell whether a model failed because of poor spatial reasoning, insufficient training data, or a gap in low-level motor control. ToyBlocks is small enough that all three can be tested separately.
What it actually measures
The framework auto-generates training demonstrations using reinforcement learning and motion planning inside NVIDIA Isaac Lab, reaching up to 24x real-time speed on a single GPU, according to Allen AI. Researchers can tune the mix, adding more flipping or pushing examples, then hold out certain task types and measure how well a trained policy transfers.
Early results from Allen AI's own evaluations are sobering: popular vision-language-action models including ManiFlow and MolmoAct 2 generalize poorly across workspace regions even after large-scale pretraining. Allen AI states that generalization "remains far from solved."
Agentic experiments with frontier models
ToyBlocks also exposes its skills through the Model Context Protocol so frontier models can control the robot via tool calls. Allen AI tested Claude Opus 5.5 on classical block-stacking tasks, reporting 30 out of 30 solved with closed-loop image observations. On the harder search-and-spell task, the agent succeeded or correctly declared infeasibility 63% of the time, with most errors involving letter rotation or look-alike confusion.
One notable result: when the agent was allowed to run its own experiments to learn how skills behave, success rates rose from 24% to 63%, surpassing descriptions written by a human expert.
The code, data generation tools, and a natural-language web interface are open and available via the project page.