Ai2 Releases ToyBlocks, an Open Simulated Robot-Arm World: It Generates Demonstrations Automatically at Up to 24x Real Time on One GPU and a Hierarchical System Scored 50.0% (15 of 30) Zero-Shot in the Real World on the Easiest Three-Block Spelling Setup With Visible, Upright Characters, but the Real Sample Is Small and the Authors Call the Findings Initial

On November 6, 2026, the Allen Institute for AI (Ai2) published ToyBlocks on the Hugging Face blog, an open simulation framework for studying how robots learn, and with it the idea of a FoundationWorld. A FoundationWorld pairs a small, controlled simulated world with tools that generate tasks and training demonstrations automatically and with metrics that follow a robot's behavior over time. ToyBlocks is the first one: a robot arm and blocks on a tabletop. Ai2 compares the idea to a petri dish, a setting where researchers can change one thing, repeat the test and trace why a robot failed. The article is Ai2's own account, and its findings are labeled initial.

The world is simple on purpose, but not trivial. Blocks carry pictures and letters on their faces, so tasks include searching, sorting and spelling words with a stack, and a robot may have to turn a block to see a hidden face. Domain randomization changes block size, table height and other properties between scenes. Skills such as pushing, flipping or rotating a block are learned in simulation with reinforcement learning and motion planning, inside NVIDIA's Isaac Lab, using privileged information such as exact block pose that a real robot would not have. A task-and-motion planner then strings skills into plans, and running plans in simulation produces recorded demonstrations with labels, without a person teleoperating the robot. Ai2 says hundreds of plans run in parallel on a single GPU, reaching up to 24 times real time. Evaluation reports more than success rates, such as execution speed, forces on objects and task progress.

The first experiments test action models. Vision-language-action models turn language and camera images into robot movements. Ai2 tested ManiFlow, π0.5 and MolmoAct 2 on tasks named Stack, Unstack, Flip and Rotate, training on part of the space and testing on new positions or stack heights. In the initial results these models generalize poorly despite large-scale pretraining, and the authors report techniques that improve it, while saying that generalization is far from solved. For high-level planning, a Molmo 2 vision-language model chooses the skill and the block, and a low-level model such as ManiFlow executes. Because the framework supplies oracles at both levels, a failure can be traced to planning or to execution.

Agentic systems were tested with Claude Opus 5.5 acting through tools exposed over the Model Context Protocol. In classical block-stacking, the agent solved all 30 text-described tasks and also all 30 when watching camera images in a closed loop. In a search-and-spell setting, the agent had to build three-letter words from up to eight blocks, and half of the scenes were infeasible because a letter was missing and the agent had to say so. It either solved the task or correctly declared it impossible in 63% of these tasks. Errors were mostly look-alike letters, 13%, and wrong letter orientation in a stack, 15%. In a separate setting, where the agent ran its own experiments and wrote its own skill descriptions, success rose from 24% with the default descriptions to 63%, against 36% with descriptions written by a human expert. These are different tasks with different denominators, and the second 63% is not the first.

Ai2 also used ToyBlocks to compare world models, which predict how a scene changes after a robot acts. For four skills, it generated 900 trajectories each, half successful and half not, fine-tuned three models and compared them. The 5 billion parameter video model Wan 2.2 predicted block motion and outcomes best, an agentic physics-simulator model called Astra tracked the gripper but struggled with contact, and a latent model called LeWM was fast but drifted over full trajectories. Adding 100 rollouts generated by the best video model raised a policy's success in an unseen region from 19% to 78%.

The sim-to-real test is the most important and the thinnest. On a Franka FR3 arm with a Robotiq gripper across thirty scene configurations, with 4 to 10 blocks seen in training, models trained only on simulation data ran zero-shot on the physical robot; simulator parameters had first been calibrated against real-robot trajectories. ManiFlow's atomic-skill success was 80.0% in simulation and 81.1% in the real world. The hierarchical Molmo 2 plus ManiFlow system, on the easiest three-block spelling configuration, with characters visible and upright, succeeded in 55.5% of simulated episodes, 111 of 200, and 50.0% of real episodes, 15 of 30. Ai2 itself says more trials are needed for statistical confidence, and an episode counts as failed if an operator stops the robot for an unsafe motion.

Our reading: ToyBlocks looks like a useful, well documented testbed, with a six-stage data pipeline, training scripts for ManiFlow, evaluation tools and a web interface where Claude or another model can control the simulated robot. Its value is a controlled setting where failures can be traced, not proof that robots now generalize. Treat all figures as Ai2's own initial results, from a block-world that is far simpler than a home or a factory, and note that the 30-episode real-world sample is small. We did not run the code or reproduce any result.