Tyler Han

🚀Stoked to announce the release of Rho, a family of open-weights 5B VLAs underlying @MSFTResearch work on Rho-alpha VLA+ (microsoft.com/en-us/research…) 🤖 Project page: microsoft.github.io/rhobotic… 📰 Tech report: arxiv.org/abs/2609.38164 ⌨️ Code: github.com/microsoft/rhoboti… 🤗 Models and data: huggingface.co/collections/m… Rho is a step in the direction of physical AI adept at adapting to new tasks and environments, motivated by a simple observation: offline supervised finetuning (SFT), the physical AI task adaptation stage "owned" by robot operators, is too data-hungry, because it tries to do too much besides learning the task per se: (1) mastering the control of the robot doing the task, and (2) exposing the model to many target task and environment variations. To address hurdle (1), we publish Rho checkpoints midtrained for @i2rt_robotics YAM Box, @Universal_Robot AI Trainer, and @FRANKAROBOTICS FR3 Duo, representative dual-arm robot setups across research labs and the industry. Rho itself is built on @Microsoft's own special Phi-series backbone and pretrained on robot data in the end-effector space, including UMI-style and @NVIDIARobotics Isaac Sim demonstrations, to facilitate its midtraining for other robots. Embodiment-midtrained Rho outperforms the strongest existing open-weights VLAs on physical robots and in simulation. And then there is hurdle (2), complicated by the tendency of offline data collection to oversample the same task situations while under-sampling those where the robot's policy actually needs help. We mitigate this by pairing with Rho with online adaptation methods, particularly FlowDAgger (microsoft.github.io/FlowDAgg…, a 🏅Spotlight🏅 at #CoRL2026), which kick in after offline SFT and use online corrections for model improvement, resulting in much more data-efficient task adaptation end-to-end. VAMs and WAMs, new zero-shot capabilities of physical AI, tactile sensing data, tactile-sensing humanoid hands, GPT-6 -- all of these hold the potential for making Rho and its successors so much more effective than we imagined just a year ago. Come to the 🤖8th Robot Learning Workshop at #NeurIPS2026 🤖 (robot-learning.ml/2026/) to geek out about these topics, and try Rho on your robots! 🙌 Congratulations to the team that has made Rho happen -- Simran Bagaria, Daphne Chen, Dean Fortier, Jianlong Fu, Michael Harrison, @TessHelleb88231, @neelsj, Dalton Moore, Galen Mullins, Michael Murray, Eduardo Salinas, and Reuben Tan! A huge shout-out to TPMs @Vivan_Amin and Ade Famoti, to @Universal_Robot for partnering with us, and to @ashleyllorens for his sponsorship and guidance in this work. Coming together from across @MSFTResearch's multiple organizations and locations in Redmond, Phoenix, New York, and Beijing, this team built more than a VLA family; it has laid a major foundation stone for @Microsoft's physical AI efforts. Truly impressive and inspiring. 🚀 #PhysicalAI #Robotics #MicrosoftResearch