Ai2 launches ToyBlocks, an open simulation world with 85,336 demonstrations for robot learning
On November 6 the Allen Institute for AI (Ai2) released ToyBlocks, an open simulation framework where robot arms learn to stack and arrange toy blocks. It is the first of a planned series Ai2 calls FoundationWorlds. The package includes simulation assets, task and demonstration generators, training and evaluation code, and two trained models, one based on Molmo 2 and one on ManiFlow. The first dataset holds 85,336 long demonstrations, about 1,600 hours of robot activity and 144 million frames.
Read the original on Ai2 on Hugging Face — ToyBlocks ↗
Why it matters
Robot research is held back by a lack of shared data and fair ways to compare methods. ToyBlocks gives labs a common, controlled world, and Ai2 reports that skills learned in it carry over to a real Franka arm: one policy scored 80.0% in simulation and 81.1% on the physical robot.
What changes
- For developers
- Researchers can download the code from GitHub and the data and models from Hugging Face, and generate new demonstrations up to 24 times faster than real time on a single GPU. Reinforcement learning support is not included yet.
- For businesses
- Companies working on robot manipulation get free training data and a benchmark, with a larger release and more skills promised within a few weeks.
Sources
- Ai2 on Hugging Face — ToyBlocks ↗ huggingface.co Primary source
Summarized by AI from official sources on November 7, 2026. Spotted a mistake? Let us know. How it works