HandUMI Collection
Open-source, hand-worn UMI variant for collecting bimanual manipulation data without a robot in the loop.
Capture layerWe make data collection cheaper, so you can focus on training robots.
HandUMI is one of the tools that makes this possible.
$110.68per capture unit
276.5 ghand-worn device
5target grippers
1-2 wkpilot turnaround
01 / Problem
Robotics is data-constrained.
Training a manipulation policy takes 300 to 1,200 high-quality demonstrations per task. Most teams still collect them by teleoperating a robot, at over $100 per usable hour.
General-purpose robots require general-purpose human data.
- Teleop demos tie up robots needed for validation.
- Teleop rigs degrade bimanual and tool-use quality.
- Raw recordings still need sync, labels, and QA.
- Robot-bound collection makes task variation expensive.
02 / Solution
Collect from humans. Keep robots for validation.
Human demonstrations are several times more sample-efficient than teleoperation, and a few hundred well-curated demos now beat thousands of mediocre ones. The winning move is better data, collected off-robot.
RoboNet captures demonstrations with HandUMI, collects with trained operators in Peru, and delivers QA-reviewed, policy-ready datasets without tying up robot arms for every run.
Dataset Pilots
Start with one task and validate signal quality, labels, and training usefulness.
Pilot datasetPeru Data Operations
Lower-cost collection nodes with standardized protocols, QA, and versioned releases.
Versioned batches03 / Why HandUMI
HandUMI is the first node in the RoboNet data network.
HandUMI is a hand-worn, open-source UMI variant for bimanual arms with parallel-jaw grippers. It mounts on the operator's thumb and index or middle fingers, opens and closes with a natural pinch, and swaps printable gripper tips for different robot targets.
- 276.5 grams
- $110.68 per unit
- Direct gripper-width sensing from a Feetech servo encoder
- Wrist-view fisheye USB camera
- PICO 4 Ultra or Meta Quest 3 tracking with wrist trackers
- Supported tips: AgileX Piper, ARX X5 2023, Dream Gripper TRLC, Trossen WidowX AI, and UMI Gripper
04 / How it works
Start with one task. Scale after the data proves useful.
- 01
Define the task.
Tell us the manipulation behavior your robot needs to improve.
Task brief - 02
Design the capture protocol.
We define objects, setup, success criteria, labels, metadata, and delivery format.
Capture protocol - 03
Collect demonstrations.
Trained operators collect demonstrations through the HandUMI workflow.
Session capture - 04
QA and package the dataset.
We review signals, labels, metadata, and package a versioned dataset.
QA report - 05
Evaluate and expand.
Your team evaluates the pilot, then we expand if the data is useful.
Versioned dataset
05 / Final vision
Full-body capture. Any embodiment.
The near-term wedge is bimanual manipulation data. The long-term RoboNet vision is a distributed network for full-body human data that can train humanoids, bimanual arms, mobile manipulators, and future robot embodiments.
- One data network, many robot embodiments.
- Embodiment-aware retargeting from human motion to robot policies.
- Standardized capture rigs, schema, validation, and long-term expansion.
06 / Use cases
Built for the manipulation tasks where generic data falls short.
Home service robots
Capture chores such as tidying, laundry handling, dishwasher loading, table clearing, and object retrieval in real home-like layouts.
Robot foundation models
Build broad human hand/object interaction datasets for embodied AI systems that need manipulation priors across tasks, objects, and environments.
Warehouse and fulfillment
Collect picking, packing, kitting, sorting, box assembly, container unloading, and long-tail SKU interaction data.
Back-of-house operations
Record repetitive packing, portioning, replenishment, bagging, staging, and delicate item handling tasks for deployable robot workers.
Lab and biotech automation
Capture vial handling, rack loading, instrument-adjacent movement, sample transfer, disposal workflows, and protocol-driven bench tasks.
Industrial workcells
Collect demonstrations for high-mix factory tasks, workstation loading, assembly support, inspection, finishing, and repetitive manual workflows.
Caregiving and assistive robotics
Record daily-living support tasks such as fetching, feeding-adjacent handling, dressing assistance, medication staging, and mobility support objects.
Infrastructure and field operations
Capture inspection, maintenance, tool handoff, part handling, and constrained-space manipulation for sites where downtime is expensive.
FAQ
Common pilot questions.
RoboNet collects task-specific manipulation datasets for robot learning teams using HandUMI and QA-reviewed workflows.
The main offer is data collection. HandUMI is the capture layer that enables the workflow.
Human demonstrations are more sample-efficient than teleoperation for the same collection time, and they capture natural dexterity that teleop rigs degrade. HandUMI keeps the robot out of the collection loop so robot hardware can stay focused on validation and embodiment-specific tuning.
The current printable tips target AgileX Piper, ARX X5 2023, Dream Gripper TRLC, Trossen WidowX AI, and UMI Gripper. Comparable parallel-jaw grippers can be supported by designing and printing a matching tip.
RoboNet combines Peru-based operations, lower facility costs, and lower-cost HandUMI hardware.
No. The model is built around cost-efficient quality: SOPs, calibration, task protocols, operator training, metadata, labels, QA review, and dataset acceptance criteria.
Yes. The recommended starting point is one task, one capture protocol, and a focused pilot dataset your team can evaluate before scaling.