# AIRoA-0 trains robot hands from handheld demonstrations

> The bimanual system learned from 16,151 hours of human demonstrations recorded without a robot. Its authors report a 78% average success rate across 10 tasks.

Oossa · 2026-11-04 · https://oossa.com/en/airoa-0-trains-robot-hands-from-handheld-demonstrations

A research team has introduced AIRoA-0, a robot system trained on 16,151 hours of people demonstrating tasks with a handheld gripper. The team says it collected no robot data during training: the robot appears only when the system is tested. The demonstrations use YUBI, an open-source gripper that people can hold to record manipulation tasks without a robot present.

From a single instruction, AIRoA-0 can tackle tasks such as sorting more than a dozen small parts into numbered cells, changing a robot vacuum’s mop pads and brush, unzipping a pencil case and folding a towel. It uses only cameras mounted on its two wrists, with no camera looking down at the whole scene. The evaluation took place at a site where none of its training data was recorded.

## How did it perform?

Across 10 evaluated tasks, the authors report an average success rate of 78%. They compare that with 27% for the strongest of four other vision-language-action models – systems that connect visual input and language instructions to robot actions – post-trained on the same task demonstrations. These are results reported by the research team; the supplied paper summary does not describe an independent evaluation.

AIRoA-0 is designed to work with different robots that can mount the YUBI gripper. The authors name the bimanual Franka, TX-G2, OpenArm and UR as examples. The system learns which parts of its wrist-camera views matter, estimates what stage of a task it has reached, and outputs actions continuously. The paper says it uses neither reinforcement learning nor on-robot corrections.

## What can others use?

The team says it is releasing the demonstration dataset, pretrained model weights and a recipe for adapting the system to new tasks. It describes the dataset as the largest open collection of handheld-gripper demonstrations with task and subtask instructions, to its knowledge. The authors say someone can record a task with YUBI and post-train the model in under four hours using one computer node with eight H200 GPUs.

That setup is different from conventional robot training, where people often collect demonstrations by controlling the robot itself. Here, the researchers’ approach is to record human hand movements first and transfer the learned behavior to a robot at test time. The paper’s results cover the 10 tasks it evaluated; they do not establish how well the system handles tasks or settings outside those tests.

## The facts

- AIRoA-0 was trained on 16,151 hours of handheld demonstrations, with no robot data collected during training.
- The authors report 78% average success across 10 tasks, compared with 27% for the strongest of four other tested models.
- The system uses two wrist cameras and was tested at a site where none of its training data was recorded.
- The team says it is releasing the dataset, pretrained weights and a post-training recipe.

## Why it matters

For robotics researchers and builders, the approach could make it possible to teach a robot by recording a person’s hands rather than collecting demonstrations on the robot itself. The authors report strong results on 10 tasks, but performance on other tasks and independent confirmation of the results remain unknown.

## Sources & references

1. [AIRoA-0: Long-Horizon, Precise, Cross-Embodiment Manipulation from Handheld Data](https://arxiv.org/abs/2611.03815) – arXiv cs.RO, 2026-11-04

Last updated: 2026-11-04