MolmoBot: Large-Scale Simulation Enables Zero-Shot Manipulation¶
Session: Speaker(s): Paper: arXiv 2603.16861 (Ai2), March 2026 Project page: allenai.github.io/MolmoBot · Blog: Ai2 blog
Paper summary (from Ai2's blog and secondary sources; full PDF not yet read)¶
- Core idea: Train manipulation policies entirely in simulation, with no real-world data and no fine-tuning, and transfer them zero-shot to real robots: an RB-Y1 mobile manipulator and a Franka FR3.
- Simulation (MolmoSpaces): More than 230k indoor scenes, 130k+ object assets, and 42M+ grasp annotations. Sources disagree on the scene count; an older page says 94k+.
- Data (MolmoBot-Data): About 1.7–1.8M expert trajectories for pick-and-place and articulated-object tasks. The count differs between paper versions.
- Policies:
- MolmoBot: a Molmo2 VLM with a flow-matching action head.
- MolmoBot-Pi0: a π0-architecture replica, for controlled comparison.
- MolmoBot-SPOC: a lightweight policy for edge deployment.
- Result (authors' claim): 79.2% on real-world tabletop pick-and-place, against 39.2% for π0.5.
- Openness: Code, training data, the data-generation pipeline, and evals are all released.
Talk notes¶
Why simulation? It is cheap, scalable, and controllable.
MolmoSpaces
- MolmoBot Engine
- Many scenes are available in MolmoSpaces.