ToyBlocks: A controlled world for robot learning and evaluation

ToyBlocks is an open simulation framework that provides a controlled world for robot learning and evaluation, centered around a robot arm manipulating blocks on a tabletop. It is designed to enable scientific investigation into robot learning by allowing researchers to vary conditions and trace failures in a repeatable environment. The framework can automatically generate training demonstrations and tasks at scale using reinforcement learning and motion planning, which helps study how models improve with more data. Researchers use ToyBlocks to investigate a range of topics, from low-level action models and high-level reasoning with vision-language models to transferring learned skills to physical robots.