MolmoAct2: Action Reasoning Models for Real-World Deployment

Summary

Presents MolmoAct2, a fully open action reasoning model built for practical deployment: MolmoER VLM backbone trained on a 3.3M-sample corpus with a specialize-then-rehearse recipe, three new datasets spanning low-to-medium cost platforms, the OpenFAST action tokenizer trained on millions of trajectories across five embodiments, and MolmoThink adaptive-depth reasoning. Evaluated across 7 simulation and real-world benchmarks.

Research notes

• Key findings: MolmoAct2 outperforms strong baselines including Pi-05 across 7 sim and real-world benchmarks; MolmoER surpasses GPT-5 and Gemini Robotics ER-1.5 across 13 embodied-reasoning benchmarks; MolmoAct2-BimanualYAM (720h) is the largest open bimanual teleoperation dataset to date
• Project page: https://allenai.org/blog/molmoact2. Released 2026-05-05.