Website · Hardware · Docs · Follow our work
RoboNet builds open-source tools and collects task-specific manipulation data for robot learning teams. We capture natural bimanual demonstrations and deliver synchronized, QA-reviewed datasets—without tying up robot arms for every run. We also work with teams to evaluate what data is useful and improve the collection and retargeting pipeline.
Our first project is HandUMI: an open-source, hand-worn interface for collecting bimanual demonstrations anywhere. It combines VR tracking, direct gripper-width sensing, and wrist-view video in a modular design for parallel-jaw grippers.
| ≈ $110 parts per capture unit |
276.5 g hand-worn device |
5 targets swappable gripper tips |
PICO + Quest tracking support |
|
01 · Define Specify the task, objects, environment, success criteria, labels, and delivery format. |
02 · Capture + QA Operators collect natural demonstrations. We synchronize signals, validate quality, and package a versioned dataset. |
03 · Evaluate + scale Test the pilot, use feedback to improve data quality and retargeting, then expand when useful. |
HandUMI records left and right wrist video, context-camera video, tracked SE(3) controller/tool poses from PICO or Meta Quest, and encoder-measured gripper aperture. The raw capture remains robot-agnostic and can be replayed or retargeted to supported embodiments later.
| Project | What it does |
|---|---|
| handumi-hw | Printable hardware, CAD, bill of materials, assembly guidance, and modular gripper tips. |
| handumi-sw | Calibration, synchronized recording, dataset validation, simulation replay, and robot retargeting. Read the docs. |
| handumi-quest-app | Meta Quest compatibility app for streaming tracked wrist poses into the HandUMI stack. |
We are starting with bimanual manipulation for home service, fulfillment, industrial workcells, labs, and assistive robotics. The long-term goal is a distributed data network that maps full-body human demonstrations to many robot embodiments.
- Working on robot data? Start with one focused task, evaluate a small sample, and decide what is useful before scaling.
- Building HandUMI? Start with the hardware guide and bill of materials.
- Contributing? Read our contribution guide, then open an issue or pull request in the relevant repository.
Collect from humans. Keep robots for validation.