Microsoft Research reports that a small set of human corrections improved its Rho model's performance on two physical robot tasks. In tests on a Franka FR3 Duo, each task received 15 additional human-supervised correction episodes.
On ten difficult starting configurations near the edge of the training distribution, test-tube assembly success rose from 9 to 21 of 30 trials, while plug insertion rose from 20 to 28. Researchers ran three trials per configuration.
Rho is a 5-billion-parameter vision-language-action (VLA) model family: it combines visual input and language instructions to generate robot actions. Microsoft released a base model and versions adapted for three robot setups, including the two-arm FR3 Duo used in these tests.
The action generator normally starts from a random vector and turns it into a sequence of robot movements. Rho's small learned policy instead selects that starting vector based on what the robot observes, using examples of corrected actions. Its main visual-language model and action generator stayed frozen; researchers supplied corrections with a SpaceMouse.
Before those robot trials, researchers adapted Rho to each task using 150 demonstrations. The 15 added episodes were human-supervised corrections during physical interaction, so the result measures improvement with continued human input.
The paper does not establish whether the gains transfer to new hardware, environments or task distributions, and it does not systematically test robustness to language instructions.
Sources
- Rho: A Foundation for Efficiently Adaptable VLA Models, Microsoft Research, September 29, 2026 (CC BY 4.0).
- Rho code repository, Microsoft Research (MIT license).
- Rho model checkpoints.
Image: Microsoft Research, Figure 1 of the Rho paper, CC BY 4.0.