Ai2 Releases Open-Source MolmoAct 2 Robotics Foundation Model
Frontier AI models
Ai2 introduced MolmoAct 2, an open-source robotics foundation model for 3D reasoning and manipulation. Alongside model weights, Ai2 released fine-tuning code, training datasets including the 700+ hour MolmoAct 2-Bimanual YAM dataset, an open FAST tokenizer, and LeRobot integration.
The Allen Institute for AI (Ai2) announced MolmoAct 2, an open robotics foundation model designed for real-world physical manipulation. Built upon Molmo 2-ER—a vision-language backbone trained on spatial reasoning—MolmoAct 2 integrates a flow-matching action expert and an adaptive-depth mechanism to process 3D spatial environments. The release includes the 700+ hour MolmoAct 2-Bimanual YAM dataset, model weights, open-source tokenizer implementations, and integration with Hugging Face's LeRobot platform.
Key takeaways
- MolmoAct 2 combines the Molmo 2-ER reasoning backbone with a flow-matching action expert and adaptive 3D depth perception.
- Ai2 open-sourced weights, code, and the MolmoAct 2-Bimanual YAM dataset featuring over 700 hours of dual-arm demonstrations.
- Inference latency was reduced from 6,700 ms in MolmoAct to 180 ms for the base model and 790 ms with adaptive depth reasoning.
- The model is integrated with Hugging Face's LeRobot framework and tested in real-world settings including laboratory biology workflows.
Source：allenai.org Read the original ↗
Prepared, checked and automatically published by AI; not individually reviewed by a human.