MolmoAct2: Action Reasoning Models for Real-world Deployment

Haoquan Fang, Jiafei Duan, Donovan Clay, Sam Wang, Shuo Liu, Weikai Huang, Xiang Fan, Wei-Chuan Tsai, Shirui Chen, Yi Ru Wang, Shanli Xing, Jaemin Cho, Jae Sung Park, Ainaz Eftekhar, Peter Sushko, Karen Farley, Angad Wadhwa, Cole Harrison, Winson Han, Ying-Chun Lee, Eli VanderBilt, Rose Hendrix, Suveen Ellawela, Lucas Ngoo, Joyce Chai, Zhongzheng Ren, Ali Farhadi, Dieter Fox, Ranjay Krishna

Open on GitHub

Type: repo
Venue: Allen Institute for AI (GitHub allenai)
Year: 2026
Source: github
Access: free
Language: en
Added: 2026-08-14T20:50:00Z
Verified: 2026-08-14T20:50:00Z

Summary

Ai2 follow-on to MolmoAct. Molmo2-ER embodied-reasoning VLM plus robot state/action modeling and a flow-matching action expert. Releases foundation checkpoints (MolmoAct2, MolmoAct2-Think with depth-token reasoning, MolmoAct2-Pretrain discrete VLA, Molmo2-ER), finetunes for DROID Franka, bimanual YAM, SO-100/101, LIBERO, and LeRobot v3 datasets (SO-100/101, DROID, BC-Z, Bridge, RT-1, MolmoAct, plus Molmo2-ER spatial/3D data). Integrated into LeRobot (molmoact2-policy). FastAPI act servers for DROID and YAM. Paper arXiv 2605.02881. Apache-2.0; research/educational use per Ai2 Responsible Use Guidelines. Hardware-safety warnings on the README.

Keywords

molmoact2 · vla · robotics · allenai · lerobot · flow-matching · droid · libero

Topics

robotics, vision-language-action, embodied reasoning

Research notes

Primary: GitHub README (Apache-2.0) plus HF allenai/MolmoAct2. Paper arXiv 2605.02881. Blog at allenai.org/blog/molmoact2. Robotics and Molmo2-ER dataset collections are substantial and not in datasets_local. Discord posted the GitHub repo.