The humanoid hardware race is moving incredibly fast. The next race may be about something far less visible: who can teach these robots the fastest?
Xingxing Wang, founder and CEO of Unitree Robotics, recently made a point that deserves more attention than it received. Robots are becoming remarkably capable. They can walk, run, recover from disturbances, manipulate objects and perform increasingly impressive demonstrations. But performing one carefully trained task is very different from being able to enter a factory and reliably learn dozens or hundreds of new ones.
This is where one of the biggest challenges of Physical AI begins. And it is not a hardware problem.
The Bottleneck Is Data — But Not the Kind You Think
A humanoid robot needs much more than videos of humans performing tasks. Video captures what a task looks like. It does not capture what a task is — not in terms a robot can act on.
What the robot actually needs are demonstrations that connect perception with real robot actions:
- how the arm moves through the workspace
- how the wrist rotates as the object is approached
- when exactly the gripper closes
- how much force the grasp requires
- how the body repositions itself around the task
- what happens when the object is slightly different from last time
That last point is where most datasets quietly fall apart. A demonstration that only ever shows the ideal case teaches the robot nothing about recovery, and recovery is most of what factory work consists of.
And ideally, these demonstrations should be collected directly on the robot that will later perform the task. Motion retargeted from a human body — or from a different robot — carries assumptions about reach, mass distribution and joint limits that do not survive the transfer.
Our Approach: Wearable Teleoperation Instead of VR
This is exactly one of the problems we are working on at Sentio Robotix with the Unitree G1.
We are developing TeleMotion, our own teleoperation system for humanoid robots. Instead of relying on a complex VR setup, we use wearable sensors to transfer human movement directly to the robot.
The concept is deliberately simple:
Human demonstration → robot motion → recorded training data.
Today we can control parts of the G1 wirelessly, and we are gradually extending the system toward both arms and full-body teleoperation.
The reason we went with wearable IMU sensors rather than VR is practical rather than ideological. A VR-based rig requires headsets, base stations, external GPU workstations and a prepared space. That is acceptable in a lab. It is a real obstacle if you want an operator on a factory floor to demonstrate twenty variations of a task before lunch. Lowering the setup cost of a demonstration directly raises how many demonstrations you can afford to collect — and volume is the whole point.
Why the Gripper Matters Just as Much
In parallel, we are developing our own lightweight grippers for the G1.
The combination is what makes this interesting. Teleoperation alone gives you motion. A gripper gives the robot the ability to interact with the environment. Together, they create something considerably more valuable: a practical data collection platform for real manipulation tasks.
Imagine a worker demonstrating:
- picking a component
- placing it into a machine
- operating a control
- moving a box
- performing a simple assembly operation
Instead of describing the task in software, the human simply performs it through the robot. The demonstration is recorded, and human knowledge gradually turns into robot training data.
There is a second, less obvious benefit. Because the demonstration runs through the actual gripper on the actual robot, the recorded force and timing data are physically real. The dataset does not need a correction layer to account for the difference between the demonstrator's hand and the robot's end effector — there is no difference, because they are the same hardware.
Teleoperation Is a Bridge, Not a Destination
Teleoperation is not the final goal. Autonomy is.
But teleoperation may be one of the most practical bridges between today's humanoid hardware and tomorrow's autonomous Physical AI. It is available now, it runs on hardware that already exists, and every hour of it produces training data that would otherwise not exist at all.
The industry has spent several years proving that humanoid robots can move. The harder and less photogenic question is how they learn to work — and the answer is going to be measured in demonstrations collected, not in demos published.
We are experimenting with exactly that on the Unitree G1.
Sentio Robotix builds practical tools for humanoid robots: lightweight grippers, wireless IMU-based teleoperation and onboard compute for the Unitree G1. sentiorobotix.com