What happened
Ai2 published ToyBlocks on November 6, 2026 via Hugging Face. ToyBlocks is the first FoundationWorld: a narrow tabletop domain (robot arm and lettered blocks) with automatic skill and demonstration generation in Isaac Lab, reaching up to 24× real-time on one GPU. The initial open data release includes 85,336 long-horizon demonstrations (about 6–12 skills each), nearly 1,600 hours of robot execution, and 144 million frames, plus simulation assets, generators, training and evaluation code, and two example models (Molmo 2 and ManiFlow). On classical block-stacking, an agent using Claude Opus 5.5 solved 30 of 30 tasks from text and 30 of 30 in a closed vision loop; on search-and-spell with up to eight blocks (half infeasible), it solved or correctly declared infeasible 63% of tasks. Zero-shot on a Franka FR3, ManiFlow atomic skills were 81.1% real vs 80.0% sim; the Molmo 2 + ManiFlow hierarchy was 50.0% real (15/30) vs 55.5% sim (111/200).
Why it matters
Robot papers often mix demos from teleoperation with opaque evals. ToyBlocks packages a regenerable sim, a named data dump, and a small real-robot transfer table so other labs can change the mixture and rerun.
Self-written skill descriptions raised agent success from 24% (default) to 63%, above 36% with expert-written descriptions.
Adding 100 rollouts from the best video world model raised a policy’s success in an unseen region from 19% to 78% in Ai2’s test.
A larger second data release is promised in a few weeks.