Ai2 has released ToyBlocks, an open simulation framework for robot learning on the Hugging Face platform. The tool is designed to help researchers study how vision-language-action models improve with more training data and how they combine reasoning with physical control. A webpage interface lets users instruct the robot using natural language via Claude or other frontier models.
ToyBlocks uses Nvidia’s Isaac Lab to generate training demonstrations and skills at up to 24x real-time speed. The framework allows researchers to generate tasks, train models, and evaluate performance in a controlled tabletop environment. It enables testing how models generalize to new spatial configurations and unseen tasks without the need for physical hardware.
Researchers can now build and test robot behaviors in a digital environment that mirrors real-world physics. This speeds up the development cycle by allowing rapid iteration on complex tasks. Developers building or buying AI for robotics should use this tool to train and evaluate models faster than in the physical world.
