On November 6, 2026, Ai2 released ToyBlocks, a controlled environment for robot learning and evaluation built around manipulating letter blocks. The release includes simulation assets, procedural task and demonstration generators, 85,336 long-horizon demonstrations (about 1,600 hours of robot execution and 144 million frames), training and evaluation code, validation sets, and two trained models based on Molmo 2 and ManiFlow. A larger second release is targeted within weeks.
Skills such as pushing and flipping are learned with reinforcement learning in Isaac Lab, and a task-and-motion planner sequences them into full demonstrations, running hundreds of plans in parallel on one GPU at up to 24 times real time. Domain randomization varies block size, friction, table height and lighting. A hierarchical agent uses Molmo 2 to choose skills and blocks and ManiFlow to execute them.
ManiFlow’s atomic skills succeeded 80.0 percent of the time in simulation and 81.1 percent on a real robot. On block-spelling tasks the combined agent succeeded 55.5 percent in simulation (111 of 200) and 50.0 percent on the real robot (15 of 30). In an agentic search-and-spell setting, learned skill descriptions raised success from 24 percent to 63 percent, versus 36 percent with expert-written descriptions. Among world models, Wan 2.2 predicted block outcomes best, and adding 100 model-generated rollouts raised a policy’s success in an unseen region from 19 to 78 percent.
Why it matters: robot learning lacks cheap, reproducible benchmarks that transfer to hardware, and a narrow but fully open world with sim and real numbers side by side makes it easier to compare policies, planners and world models on equal terms.
What it does not show: real-robot trial counts are small (30 trials for spelling); the scope is deliberately narrow; action-model generalization remains far from solved; frontier models still misread letters; and RL training is not yet integrated in the package.