Ai2 ToyBlocks: open world for robot learning

What changed?

Allen Institute for AI (Ai2) launched FoundationWorlds: controlled simulated worlds for robotics research with automatic task and demo generation plus evaluation metrics. The first world is ToyBlocks: a robot arm and tabletop blocks with letters or pictures on their faces, built as an open simulation framework.

Auto demos and VLA tests

The stack generates demonstrations without teleoperation. In Isaac Lab it combines RL, motion planning and TAMP, running up to 24× realtime on one GPU. Researchers studied spatial generalization of VLA models (ManiFlow, π0.5, MolmoAct 2). Early results show poor generalization; the framework is meant to help improve those techniques systematically.

Hierarchy, agents and world models

For long-horizon work, Molmo 2 handles high-level planning and ManiFlow the low-level motion, for example spelling tasks. With Claude-backed MCP tools, search-and-spell reaches about 63 percent solve-or-correctly-infeasible. Agent-written skill descriptions lift success from 24 percent to 63 percent.

Among world models, Wan 2.2 led on pose and success prediction. With 100 model rollouts, policy success in an unseen region rose from 19 percent to 78 percent.

Sim-to-real and the release

Zero-shot sim-to-real on a Franka FR3: ManiFlow skills scored about 80 percent in sim and 81.1 percent in reality. Molmo 2 plus ManiFlow reached 55.5 percent in sim (111/200) versus 50 percent in reality (15/30).

Ai2 released assets, generators, demos, two trained models, plus training and eval code. The initial set has 85,336 long-horizon demos (about 6-12 skills each), roughly 1,600 hours and 144 million frames. A larger release is due in a few weeks.