Microsoft

Microsoft is a global technology company. Its entry into embodied AI foundation models is the Rho family (arXiv 2609.38164, 2026-09-29): a family of open-weights vision-language-action models for bimanual manipulation whose backbone is distilled from Microsoft's Phi-family VLM and coupled to a flow-matching action expert. The Rho paper states that the base model and embodiment-specific checkpoints (Rho-YAM-Box, Rho-UR-AI-Trainer, Rho-FR3-Duo) are released through the official microsoft/rho HuggingFace collection, described as "a family of physical AI models for robotic manipulation". Rho's studies were performed in the RoboTwin, LIBERO, and RoboEval simulation environments and on physical dual-arm platforms. Microsoft maintains one of the largest HuggingFace organizations and a broad open-source research portfolio; Rho marks its direct release of open-weights robot manipulation foundation models.