EN ▾
SNAPE REP RT Front page →
ConfirmedRobotics06 Nov, 14:00 UTC

Ai2 open-sources ToyBlocks for controlled robot-learning experiments

ToyBlocks simulated robot arm manipulating lettered blocks on a tabletop
Ai2 / Hugging Face

Hugging Face blog: FoundationWorld tabletop sim with auto demos, 85,336 long-horizon demonstrations, Molmo 2 + ManiFlow models, and reported 50% zero-shot sim-to-real spelling success on a Franka arm.

Read at Hugging Face (Ai2) Share

Why it matters

ToyBlocks is Ai2’s first FoundationWorld: a narrow block-and-arm domain meant to make robot-learning failures traceable by controlling training data and evals. The release includes generators, assets, training and evaluation code, and two trained models. Early VLA generalization tests were poor until fine-tuning tricks; hierarchical Molmo 2 + ManiFlow transferred to a physical Franka with similar skill rates to simulation. Watch the promised larger dataset and integrated RL follow-up.

Key facts

  • Published on Hugging Face blog 2026-11-06T14:00:48Z
  • 85,336 long-horizon demonstrations (~1,600 hours, 144M frames)
  • Molmo2-4B-tb planner + ManiFlow-tb low-level control
  • Hierarchical system 50% success on physical Franka spelling tasks (15/30)
  • ManiFlow atomic skills ~80% sim and ~81% real

Sources

More on Robotics
More from Snape ReportFront page →