China’s National University of Defense Technology, in collaboration with Midea Group, has introduced an innovative system called “HumanoidExo”. This technology features a wearable suit that records a person’s full-body motion, allowing humanoid robots to learn directly from human movements.
Humanoids typically lose balance because their training data often comes from videos or computer simulations. HumanoidExo addresses this issue by capturing precise, real-time human motion through wearable sensors.
“A significant bottleneck in humanoid policy learning is the acquisition of large-scale, diverse datasets, as collecting reliable real-world data remains both difficult and cost-prohibitive,” the researchers noted in their paper.
How the HumanoidExo system works
The wearable suit in the “HumanoidExo” system tracks seven arm joints and synchronizes them with the robot’s movements. It also uses wrist-mounted motion sensors and a LiDAR scanner on the back to measure the wearer’s height and motion changes.
The system operates with a dual-AI model known as HumanoidExo-VLA. One component, the Vision-Language-Action model, interprets human tasks. The other, a reinforcement learning controller, keeps the robot balanced while it learns to walk and perform complex tasks.
Experimenting with the Unitree G1 humanoid
In experiments, the researchers used both teleoperation and exoskeleton-recorded sessions to train the “Unitree G1 humanoid robot”. The combination of 200 training sessions increased the robot’s success rate in pick-and-place tasks from just 5% to nearly 80%.
Surprisingly, the robot also learned to walk—even though the training dataset did not contain direct walking instructions. When physically pushed off balance, it managed to walk back to its original position and complete its task.
Global context and industry implications
This breakthrough arrives during an international surge in humanoid robot development. Companies such as NVIDIA, Google DeepMind, Tesla, and Figure AI are racing to create robots capable of learning from large-scale human data.
“We’re seeing humanoid robots everywhere—in the U.S., China, from Tesla, from Figure AI,” said Wandercraft CEO Matthieu Masselin. “Once we began getting more requests and people pulled us into that market, it made sense to develop, alongside our exoskeleton, a free and autonomous humanoid robot that relies on the same technology.”
What this means for the future
The “HumanoidExo” project represents a major step toward bridging the gap between human movement and robotic precision. By capturing authentic physical motion rather than relying on simulation data, researchers can accelerate humanoid learning while improving safety and coordination.
As robotics becomes increasingly intertwined with AI, innovations like HumanoidExo could transform how machines learn and move—bringing us closer to truly human-like robots that adapt and operate independently.
Source: https://arxiv.org/pdf/2510.03022
