ToyBlocks: Controlled simulation for robot learning research

Hugging Face has introduced ToyBlocks, an open simulation framework for studying robot learning in a controlled environment. Built on Nvidia's Isaac Lab, the system uses automated task-and-motion planning to generate training data at scale, reaching up to 24 times real-time performance. The framework pairs high-level reasoning models with low-level action policies to test generalization and spatial reasoning in a reproducible setting. This allows researchers to isolate variables like training data composition and model architecture without the noise of real-world physical constraints.