AI2's ToyBlocks: Sim-Trained Robot Stacks Words on Real Hardware, 15/30 Attempts
allen_ai · x · 2026-11-06
Allen AI reports ToyBlocks' key result: Molmo 2 + ManiFlow, trained only in the ToyBlocks virtual world, stacked lettered blocks to spell a word in 15/30 attempts on a physical robot, versus 111/200 in simulation. Simulation-trained models can transfer to real hardware, though more trials are needed. In earlier tests, ManiFlow, π0.5, and MolmoAct 2 all generalized poorly to new spatial configurations.
Related event: Sim-Trained Robot Succeeds in Real-World Block Tasks; Ai2 Outlines Roadmap(2 posts)→
More from Embodied
- RoboColosseum First Results: 5 Robot Models and 3 Frontier AI Agents Tested Across 124 Real-World Sites — DJiafei · 2026-11-06
- First RoboColosseum results: 5 robot models and 3 frontier agents tested across 124 real-world sites — DJiafei · 2026-11-06
- Martin Casado amazed: single-image 3D splat of Atlas, consistent novel views, no ghosting — drfeifei · 2026-11-06
- Repurposing a Logi 9-key macro pad as physical YES/NO/CONTINUE buttons for coding agents — flngr · 2026-11-06
- Community mod brings real DLSS Frame Gen to RTX 3090, matching 5090 frame-for-frame — kingslyj · 2026-11-06
- ToyBlocks Offers a Tool Interface to Drive Robots with Frontier Models via Natural Language — allen_ai · 2026-11-06