Robot demos are impressive and hard to learn from: a video of a robot folding laundry rarely tells you what it learned or why it fails. Ai2, the Allen Institute for AI, has released ToyBlocks, a deliberately small simulated world meant to work like a petri dish for robot learning, where researchers can change one thing, rerun a test and trace a failure.
The world is a robot arm and lettered or pictured blocks on a table. Tasks include pushing, flipping, stacking and searching, since a robot may have to turn a block to see a hidden face. Skills are learned with reinforcement learning and motion planning in Nvidia's Isaac Lab, and a planner chains them into long tasks, producing demonstrations automatically rather than through hours of human teleoperation. Running hundreds of plans in parallel on one GPU reaches up to 24 times real time. The first release has 85,336 long demonstrations, nearly 1,600 hours of robot execution and 144 million frames, plus two trained models, training code and evaluation sets.
The early findings are instructive. Popular action models, including pi0.5, ManiFlow and MolmoAct 2, generalised poorly at first when blocks appeared in regions or stack heights they had not trained on; Ai2 says the problem remains far from solved. Transfer to hardware looked better: ManiFlow's single-skill success was 80.0% in simulation and 81.1% on a real Franka arm, and a combined system using Molmo 2 for planning reached 55.5% of simulated spelling tasks and 50.0% of real ones, 15 of 30. Ai2 also ran Claude Opus 5.5 as an agent through MCP tools. It solved all 30 classic block-stacking puzzles, and on a harder task, finding letters to spell a word or declaring it impossible, it succeeded 63% of the time, tripped mostly by look-alike letters and sideways blocks. When the agent ran its own experiments and wrote its own skill descriptions, success rose from 24% to 63%, beating 36% with an expert's descriptions.
For readers following our coverage of robot models and of Claude Opus 5.5, the useful part is the controlled comparison. A shared, open world where results are known to carry over to a real arm lets labs test claims on equal terms instead of through curated videos.
The limits are built into the design. Blocks on a table are a narrow slice of physical work, and good transfer here does not guarantee transfer for cloth, liquids or cluttered kitchens. The real-robot trials are small, which Ai2 itself says need more runs for statistical confidence. This is a release write-up by the developer, not a peer-reviewed study.