The Allen Institute for AI (Ai2) released ToyBlocks on November 6, an open simulation framework built around a robot arm and blocks on a tabletop. Ai2 calls it its first "FoundationWorld," designed to generate robot training data and evaluations at scale without human teleoperation.
The framework uses reinforcement learning and motion planning to learn individual skills, trained in Isaac Lab, and a task-and-motion planner to chain those skills into longer demonstrations. It randomizes properties such as block size, table height and lighting, and measures more than success rate, including execution speed, forces on objects and task progress. A web interface lets users connect Claude or another frontier model to instruct the simulated robot in natural language.
The initial release contains 85,336 long-horizon demonstrations, each made up of about six to 12 skill demonstrations, amounting to nearly 1,600 hours of robot execution data and 144 million frames. Ai2 says plans run in parallel on a single GPU at up to 24 times real time. The release includes simulation assets, procedural task and demonstration generators, the generated data, two trained models (one based on Molmo 2 and one on ManiFlow), training code and evaluation code. Ai2 says a second release with a larger dataset, more skills and fuller benchmarking is targeted within weeks.
Ai2 reports that ManiFlow atomic skills succeeded 80.0% of the time in simulation and 81.1% on a real Franka FR3 arm, suggesting the simulator transfers to hardware for basic skills. On block-spelling tasks, a Molmo 2 plus ManiFlow system scored 55.5% in simulation (111 of 200) and 50.0% in the real world (15 of 30), which the authors note needs more trials for statistical confidence. In a world-model experiment, adding 100 model-generated rollouts raised a policy's success in an unseen region from 19% to 78%.
For robotics researchers, the value is a controllable, fully open pipeline: training mixtures can be generated and varied on demand instead of collected by hand, and policies can be compared on the same tasks with shared metrics. Ai2's blog post does not state a license; it points to the code repository for details.