ToyBlocks is an open simulation framework built around a robot arm and blocks on a tabletop, and Ai2 presents it as its first FoundationWorld: a controlled simulated setting paired with tools for generating tasks and demonstrations and measuring robot behavior. Its deliberately constrained domain includes pushing, flipping, stacking, searching, sorting and arranging blocks; varied block faces and randomized scene properties support tests of how models generalize across tasks and geometry.
ToyBlocks automatically generates demonstrations in simulation without time-consuming teleoperation, lets researchers vary training experiences and supports evaluation on tasks withheld from training. It uses reinforcement learning and motion planning to generate short-horizon skills in Isaac Lab, while a task-and-motion planner selects and sequences skills and plans the movements between them; hundreds of plans can run in parallel on one GPU, reaching up to 24 times real-time performance for data generation. Models can be trained to map raw sensor observations to actions, so they must estimate a block’s position and orientation rather than receive the simulator’s direct state information; one setup uses Molmo 2 to select high-level skills and ManiFlow to control the simulated arm.
Researchers can vary training mixtures and evaluate held-out tasks using measures such as success, execution speed, forces on objects and task progress; a web interface also lets users connect Claude or another frontier model for natural-language interaction.
ToyBlocks supports tests of generalization across skill sequences, block letters or colors, block counts and workspace positions; initial results for ManiFlow, π0.5 and MolmoAct 2 report poor generalization, while also describing techniques tested to improve it. Long-horizon tasks and high- and low-level reference oracles can help researchers distinguish planning failures from execution failures. For example, a classical-planning system using Claude Opus 5.5 solved all 30 text-described block-stacking tasks and all 30 tasks in a closed loop using camera observations. In a separate search-and-spell experiment, the agent solved or correctly declared infeasible 63% of tasks; reported errors included confusion between look-alike letters (13%) and incorrect letter orientation (15%). Agent-written skill descriptions achieved 63% success, compared with 24% for default descriptions and 36% for descriptions written by human experts.
ToyBlocks also provides a controlled setting for testing action-conditioned world models: in the reported tests, Wan 2.2 predicted block motion and task outcomes most accurately, while Astra struggled with contact and LeWM drifted over full trajectories despite better performance on short interactions. In another experiment, adding 100 model-generated trajectories from the best video-generation model raised a policy’s success in an unseen region from 19% to 78%. For sim-to-real tests, policies trained exclusively on ToyBlocks data were evaluated with a Franka FR3 arm and Robotiq gripper across 30 scene configurations. ManiFlow’s atomic-skill success rates were 80.0% in simulation and 81.1% in the real world; the hierarchical Molmo 2–ManiFlow system succeeded in 111 of 200 simulated episodes (55.5%) and 15 of 30 real-world episodes (50.0%). The source says further trials are needed to establish statistical confidence across tasks.
The transfer workflow tunes simulated robot parameters, randomizes physical and visual properties, and uses front-facing and wrist-mounted cameras; for policies using depth images, FastFoundationStereo generates depth maps from paired RGB images.
The release includes simulation assets, procedural task and demonstration generators, generated demonstrations, two trained models based on Molmo 2 and ManiFlow, training and evaluation code, and validation sets. Its initial data release contains 85,336 long-horizon demonstrations, nearly 1,600 hours of robot execution data and 144 million frames. The repository describes ToyBlocks as based on IsaacSim and IsaacLab, with datasets in LeRobot format and workflows for policy training and closed-loop, GPU-parallel evaluation. ToyBlocks builds on earlier open robot-learning work named MolmoAct 2, MolmoBot and MolmoSpaces. The repository states that ToyBlocks is licensed under Apache License 2.0; certain submodules and downloaded assets have separate licenses.