Paper summary
Paper summary: MolmoAct2: Action Reasoning Models for Real-world Deployment
MolmoAct2 is a fully open action-reasoning model for practical robot deployment, with a new embodied-reasoning VLM backbone, open datasets including 720 hours of bimanual teleoperation, an open action tokenizer, and a lower-latency reasoning variant.
- It's open across the stack: backbone, datasets, action tokenizer and model.
- MolmoThink re-predicts depth tokens only for scene regions that change, which keeps geometric grounding at much lower latency.