IO-AI Tech, a startup based 45 minutes north of downtown Shenzhen, is paying workers to wear VR headsets and motion-tracking gear so they can pilot humanoid robots on factory floors and inside convenience stores. The robots do useful work — stocking shelves, picking items from bins, folding laundry — while IO-AI Tech harvests the teleoperation logs as training data for future autonomy. The company is also building the software layer that lets a single operator drive dozens of different humanoid platforms.
The scale of the data-collection effort is the point. A custom motion-tracking glove demonstrated to a visiting reporter mapped one person's finger movements onto 50 robotic digits across 10 humanoid hands from different manufacturers simultaneously, with haptic feedback in both directions. That cross-platform compatibility matters in China, where dozens of different humanoids and robot hands are already on the market and no single form factor has won.
The commercial pilots are concrete. A Chinese convenience store chain is testing the rig for shelf-stocking, with operators wearing VR headsets and using a pair of grippers to pick boxes of medication off shelves. In a separate demo, workers wearing body-tracking sensors used Unitree humanoids to mirror their movements inside a mocked-up apartment — including removing a shirt from a hanger and folding it through the robot's eye-level cameras.
Key facts
- 01IO-AI Tech, based 45 minutes north of downtown Shenzhen, pays workers to teleoperate humanoid robots using VR headsets and motion-tracking gear.
- 02A single custom glove can drive 50 robotic digits across 10 different humanoid hands simultaneously.
- 03The startup's software translates one operator's movements onto dozens of different humanoids and robot hands sold in China.
- 04Jack Sewing Machines is working with IO-AI Tech to train two-armed robots to iron shirts on existing clothing production lines.
- 05A Chinese convenience store chain is piloting the system for shelf-stocking, using VR headsets and grippers.
IO-AI Tech's algorithms have to do more than copy human motion frame-for-frame. A person and a robot rarely share the same shape, size, or weight distribution, so the system blends operator input with onboard autonomy to keep the machine balanced. That hybrid control loop is the company's core technical bet: teleoperation as a scaffold, not an endpoint.
Cofounder Si Chin says the approach mirrors how self-driving cars were developed — incrementally, in narrower domains, with task-specific data rather than a single general model trained on everything at once. She told Wired that focused training data is what unlocks each new deployment, and that vocational schools in China are now training students on robot teleoperation as a job category in its own right.
Shenzhen is the natural base for this kind of work. The city's hardware manufacturing density means IO-AI Tech can iterate on prototypes quickly and partner with local industrial customers looking to automate specific bottlenecks on their existing production lines. The startup's pitch to those manufacturers is that they don't need to wait for fully autonomous robots — they can deploy teleoperated ones now and let the system gradually take over.
One of those partners is Jack Sewing Machines, which makes clothing manufacturing equipment and is working with IO-AI Tech to train two-armed robots to iron shirts. The executive's framing is telling: the robots are designed to slot into lines currently staffed by human hands, not to replace whole factories. That incrementalism is how teleoperation-first deployment differs from the more ambitious end-to-end humanoid pitch coming out of US labs.
“These robots could fit onto an existing production line and automate work currently done by hand”— Jack Sewing Machines executive, Jack Sewing Machines
The bet runs against a competing thesis in the field. Some roboticists argue that piling vast quantities of teleoperation data into a single large model will eventually produce a general-purpose robot brain, the way scaling language data produced general-purpose chatbots. Chin doesn't dismiss that path, but she's building the company on the assumption that domain-specific data sets, gathered one workflow at a time, will get robots into paying jobs faster.
The harder questions are about whether teleoperation economics actually work. A human operator wearing a VR rig to stock a convenience store shelf is not cheaper than a human stocking the shelf directly — the value is in the data flywheel and the eventual handoff to autonomy. If that handoff takes longer than expected, or if a competing approach using simulation-trained policies (the direction Nvidia's ENPIRE system is pushing) gets to deployment first, the teleoperation labor cost becomes hard to justify.
China's lead in humanoid hardware — Unitree's machines being the most visible example — combined with IO-AI Tech's data-collection layer points to a different industrialization path than the one Western robotics startups are pursuing. The Western pitch tends to be a single capable humanoid sold as a product. The Shenzhen pitch is a teleoperation marketplace that absorbs whatever humanoid the customer already bought and turns operator hours into training tokens. Whichever model produces the first profitable factory deployment will set the template for how humanoids actually reach scale.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
