Portuguese 
Spanish 
Why folding clothes is still difficult for robots
In Shenzhen, in southern China, workers are using virtual reality glasses, manual controls, and body sensors to remotely command humanoid robots in tasks that simulate store, factory, and domestic routines.
Each movement made by the operator turns into data to train physical artificial intelligence systems, an area that seeks to teach machines to handle objects, surfaces, balance, force, and real-world situations.
The operation was described by WIRED in a report published on June 17, 2026, after a visit to IO-AI Tech, a Chinese startup dedicated to teleoperation and data collection systems for humanoid robots.
-
12-Year-Old Math Prodigy from Brazil Wins Two Gold Medals and Aims for Elite IMPA Tech
-
Magnet Weighing Over 360 Kilograms, Designed for Cardiac Surgery Machine and Valued at $100,000, Retrieves Sawed-Off Shotgun with Scratched Serial Number from Sea During Magnetic Fishing
-
Largest Mechanical Dredger in the Western Hemisphere Removes Nearly 80 Cubic Meters of Mud Per Cycle, Measures 77 Meters, and Operates Day and Night to Prevent Giant Ships from Getting Stuck in the Houston Ship Channel
-
California Opens Floodgates to Restore Wetlands and Create Natural Flood Protection in the Central Valley
In this model, the worker not only performs a task through the machine but also helps form databases used to improve the autonomy of future robotic systems.
How teleoperation turns human gestures into data
To control the robot, the operator wears a headset, holds controls, or uses tracking gloves, beginning to see the environment from the machine’s perspective while moving arms, hands, and legs.
From this first-person view, they can pick up items, organize shelves, manipulate objects, and repeat movements in scenarios set up precisely to bring the training closer to real situations.
According to IO-AI Tech, the TeleXperience platform allows controlling different robot configurations and collecting real high-precision data to train humanoids, robotic arms, and mechanical hands.
More than remote control, the technology records information about vision, force, reach, posture, and object reaction, elements necessary for artificial intelligence to learn from physical experiences.
In a demonstration reported by WIRED, operators controlled Unitree humanoid robots inside a space that simulated an apartment, with furniture, hangers, and everyday objects.
While a person walked and moved their body, the machine followed the gestures and performed actions like taking a piece of clothing off a hanger and folding it.
Why folding clothes is still difficult for robots
Simple tasks for humans, like folding a t-shirt or picking up a box from a shelf, require a sequence of adjustments that robots still need to learn precisely.
In these activities, the machine must calculate force, interpret distance, maintain balance, correct posture, and adapt movement when the object changes position or reacts unexpectedly.
Flexible objects make the process even more complex, as fabrics fold, slip, create creases, and change shape with each contact with mechanical hands.
For this reason, the presence of the operator helps reduce errors during training, as they correct movements in real-time and guide the machine in difficult-to-predict situations.
The difference between the human body and the robotic body also requires constant adjustments, because height, weight, reach, and balance do not always coincide between the commander and the executor.
According to WIRED, the startup’s systems need to combine human command with some degree of autonomy to adapt gestures to robots of different shapes, sizes, and weights.
Physical artificial intelligence depends on real experiences
Unlike digital models trained with large volumes of texts, images, and videos, humanoid robots need data produced during direct interactions with the environment.
For a physical machine, visually recognizing a box or a t-shirt is not enough, because the task also requires knowing where to touch, how much force to apply, and how to react to contact.
It is at this point that the worker takes on a different role from traditional industrial operation, as their gesture becomes part of the system’s learning process.
In addition to guiding the machine in an immediate task, the operator produces records that can be reused in the development of more autonomous robots less dependent on human intervention.
IO-AI Tech also presents SenseXperience, focused on capturing human data in the real world, and EmbodiFlow, aimed at collecting, annotating, visualizing, and post-processing multimodal data.
These tools indicate that the advancement of humanoids depends not only on motors and sensors but on large volumes of organized physical examples for embedded artificial intelligence training.
Shenzhen concentrates robotics, sensors, and prototypes
The choice of Shenzhen favors this type of experiment because the city brings together manufacturers, component suppliers, hardware companies, and startups capable of quickly adjusting prototypes.
In this industrial ecosystem, parts, sensors, and structures can be modified in short cycles, which facilitates the integration of teleoperation with different robot models.
WIRED reports that IO-AI Tech is working with local manufacturers interested in automating manual steps, including activities related to the clothing sector and handling parts.
Among the examples cited is Jack Sewing Machines, a Chinese sewing equipment company involved in a project to train two-armed robots in tasks such as ironing shirts.
In retail, the technology was also tested with a Chinese convenience store chain, in an operation where a person used a headset and controls to pick up boxes of medication.
This type of demonstration combines practical execution and data collection, allowing the recording of object positions, arm reach, environmental vision, and movements necessary to complete the task.
Robots still depend on human operators
Despite the goal of increasing autonomy, teleoperation shows that many humanoid robots still need human supervision to operate in complex and unpredictable environments.
Factories, stores, and homes were designed for human bodies, with corridors, shelves, furniture, and tools that require mobility, perception, and constant adaptation.
For this reason, humanoids attract the interest of companies seeking machines capable of working in existing spaces without requiring a complete reconstruction of the surrounding environment.
Even so, the path to autonomy involves millions of captured gestures, corrected and converted into data, with workers serving as a bridge between the current machine and future systems.
At this stage of physical robotics, the value of work is not only in the task performed at the moment but also in the knowledge incorporated into the records that help train new generations of machines.
