MolmoBOT: 1.7M Simulated Trajectories Enable Real Robot Skills With Zero Real Data
chris_j_paxton · x · 2026-10-23
MolmoBOT demonstrates that generating sufficiently diverse simulation data can bypass the sim-to-real transfer problem entirely. The open-source project has three parts: MolmoBot-Engine, which produces diverse pick-and-place demonstrations across many simulated environments; MolmoBot-Data, a dataset of 1.7 million expert trajectories for object manipulation; and several VLA policies trained on it that achieve strong performance on real-world pick-and-place tasks with zero real robot data. Chris Paxton highlights it as a path to unlocking the missing massive scale of robotics data.
More from Embodied
- A $2,490 humanoid and a hosted robot-model API land in the same week — AIFlow_ML · 2026-10-24
- iPhone Duo's fold transition bends the focal plane, not the content — emax · 2026-10-23
- Microduck: why this cute robot toy matters more than it looks — TinfoilTricorn · 2026-10-23
- Rust SLAM system visloc-rs debuts 4th on LaMAria benchmark, top open-source entry — rsasaki0109 · 2026-10-23
- AGIBOT's A3 Ultra humanoid robot demonstrates roller-skating mobility — TinfoilTricorn · 2026-10-23
- Xaira details X-Automate: X-Scientist, X-Bot and mobile robots link protein design to testing — BoWang87 · 2026-10-23