Cohere presents MolmoAct2, an open robotics foundation model

The model uses a three-stage training pipeline with open datasets and enables zero-shot reasoning for real-world deployment. It aims to address limitations of current vision-language-action models in robotics.