Rho
Microsoft's open 5B VLA, pre-adapted to three common dual-arm robots
License
MIT
Overview
Rho is a family of open-weight vision-language-action models for two-armed robots from Microsoft Research, released in September 2026 under the MIT licence. Each model is about 5 billion parameters: a 4.68B Phi-family vision-language model with extra training on physical grounding, joined to a 542M flow-matching action expert that outputs about a second of end-effector motion at a time. The idea is to adapt in two steps. Rho-base, pretrained on more than 3,200 hours of mostly dual-arm robot data, is first midtrained on about 300 hours from one specific robot. The released checkpoints have already done that for the I²RT YAM Box, the UR AI Trainer and the Franka FR3 Duo. Task finetuning then needs fewer demonstrations. The code finetunes on LeRobot datasets, evaluates on LIBERO and RoboEval, and serves a policy over a websocket. All weights and datasets are on Hugging Face, and the method is described in the technical report.
Most open VLAs are pretrained broadly and then finetuned straight onto a task, so the demonstrations have to teach the robot and the task at once. Rho's premise is to learn the robot first. Between pretraining and task finetuning sits an embodiment midtraining stage: two epochs over roughly 280 to 300 hours of one robot's data across dozens of task families. It is not meant to master any task, only to absorb that robot's cameras, kinematics, workspace and control conventions. The authors release that stage already done for three widely used dual-arm platforms, so a lab that owns one starts from a model that already knows its hardware.
The evidence for the idea is in controlled comparisons rather than a single headline. In the RoboTwin 2.0 simulator, the midtrained model reached the same success with half the finetuning demonstrations as Rho-base, with the biggest gap at the smallest data budget. On a physical FR3 Duo, the same 150 plug-insertion demonstrations gave 60% success from Rho-base, 76.7% after one epoch of midtraining and 86.7% from the released two-epoch checkpoint. On the YAM Box BusyBox benchmark, with about 2,000 demonstrations, the finetuned Rho-YAM-Box reached 90% overall, level with π0.5 and ahead of GR00T N1.7 and MolmoAct2.
The release also includes an online stage. A small latent policy inside the model learns from human corrective takeovers while everything else stays frozen. In the report, 15 corrected episodes raised the hardest test-tube assembly configurations on the FR3 Duo from 30% to 70%. The state and action interface is generic: up to 32 dimensions per step, masked when unused. So the same model can drive joint or end-effector control, deltas or absolute targets, and one arm or two.
Why it matters
The expensive part of using a VLA is not downloading it but adapting it to your robot and your task, which usually means hundreds of demonstrations per task. Rho moves the robot-specific part of that cost into a published checkpoint for three platforms many labs already own, under the MIT licence. The report backs the claim with controlled comparisons against π0.5, GR00T N1.7 and MolmoAct2 on physical robots. It also gives open VLAs a fourth serious family, from a major lab, alongside OpenPI, GR00T and MolmoAct.
Problem it solves
Finetuning a general VLA onto a new task forces the demonstrations to teach both the robot and the task. Rho separates the two: the robot is learned once, in a midtrained checkpoint that is released for three dual-arm platforms, so task finetuning starts from a model that already knows the hardware and needs fewer demonstrations. A lightweight online stage then keeps improving a deployed policy from a handful of human corrections.
Key takeaways
- Open-weight 5B VLA: a 4.68B Phi-family VLM with physical-grounding training plus a 542M flow-matching action expert
- Rho-base is pretrained on more than 3,200 hours of mostly bimanual robot data, then midtrained into variants for YAM Box, UR AI Trainer and FR3 Duo
- In RoboTwin 2.0 the midtrained model matched Rho-base with half the finetuning demonstrations
- Finetuned Rho-YAM-Box reached 90% on BusyBox, level with π0.5 and ahead of GR00T N1.7 and MolmoAct2
- Online adaptation from 15 corrected episodes lifted the hardest FR3 Duo test-tube configurations from 30% to 70%
- MIT-licensed code, weights and finetuning datasets; finetunes on LeRobot datasets and serves over a websocket
Who it is for
- Labs with an I²RT YAM Box, UR AI Trainer or Franka FR3 Duo who want a pre-adapted starting policy
- Teams comparing open VLAs on bimanual manipulation who need a strong MIT-licensed baseline
- Researchers studying data-efficient adaptation, midtraining and learning from human corrections
- LeRobot users who want to finetune a 5B VLA on their own datasets
Limitations
- The 3,200-hour pretraining mixture and the per-robot midtraining data are not released; only the checkpoints, the BusyBox set and the LIBERO and RoboEval finetuning data are
- Pre-adapted variants exist only for three dual-arm platforms; any other robot starts from Rho-base and needs its own midtraining or more task data
- Comparisons are the authors' own, on their robots and protocols; the project page itself says three platforms and a handful of benchmarks do not establish general performance
- Task policies still need roughly 100 to 300 demonstrations on a physical robot; midtraining reduces that, it does not remove it
- Training and inference need a CUDA GPU and FlashAttention 2 built against a CUDA toolkit; the Docker route needs about 100 GB of build space
- Released at the end of September 2026, so there is no outside track record yet