ToyBlocks

ToyBlocks is Ai2’s first FoundationWorld: an open, controlled robotics simulation for a tabletop arm manipulating blocks (push, flip, stack), built on NVIDIA Isaac Sim / Isaac Lab.

Highlights:

• Automatic skill and long-horizon demonstration generation without teleoperation;
• Initial release ~85,336 long-horizon demos, ~1,600 hours, 144M frames (LeRobot format);
• Training (e.g. ManiFlow) and closed-loop eval; swap either hierarchy level for an oracle to isolate failures;
• Goals in PDDL with automatic success / partial-progress metrics.

Blog: https://huggingface.co/blog/allenai/toyblocks Repo: https://github.com/allenai/ToyBlocks

Why it matters

Robot learning is often bottlenecked by expensive demos and hard-to-trace failures. ToyBlocks co-locates data generation, training, and evaluation in one controlled world so teams can separate planning vs control failures and study sim→real transfer.

Caveats

• FoundationWorld is Ai2’s series name; ToyBlocks is the first instance;
• Expect Isaac Sim/Lab and NVIDIA GPU setup;
• Disambiguate from toy brands — use allenai/ToyBlocks.

1. Read https://huggingface.co/blog/allenai/toyblocks ;
2. Clone https://github.com/allenai/ToyBlocks and set up Isaac Sim/Lab per docs;
3. Train with allenai/toyblocks-data or self-generated demos;
4. Evaluate closed-loop in sim; use oracles to isolate hierarchy failures.