Экзоскелет HumanoidExo помогает роботам овладеть человеческими движениями с минимальными данными Translation: HumanoidExo Exoskeleton Helps Robots Master Human Movements with Minimal Data

Researchers from the National University of Defense Technology in China, in collaboration with Midea Group, have unveiled the HumanoidExo system—an advanced wearable suit designed to capture human movements and convert them into structured data for robot training.

These experts aim to tackle one of the most challenging issues in robotics: teaching humanoid robots to move like humans without the necessity of gathering thousands of expensive demonstrations.

During trials, the newly trained humanoid Unitree G1 was able to perform complex maneuvers and even walk after only a few examples.

«A significant bottleneck in robot training is the acquisition of large, diverse datasets since gathering reliable real-world information remains a complex and costly endeavor,» the researchers noted.

Humanoid robots often struggle to replicate human movements fully, as their training is primarily based on video footage and simulations. HumanoidExo addresses this gap by capturing real human motions.

The suit directly correlates the seven joints of a human arm with the robot’s configuration. It employs inertial sensors at the wrists and a LiDAR module on the back for tracking the position and height of the user.

The data is fed into a two-level AI system known as HumanoidExo-VLA, which consists of a Vision-Language-Action model that interprets tasks and a reinforcement learning controller to maintain balance and stability in movements.

According to the researchers, Unitree G1 was trained using just five teleoperated demonstrations and 195 sessions recorded using the exoskeleton. This hybrid approach elevated its task success rate for object transportation from 5% to approximately 80%, matching the performance of models trained on 200 demonstrations.

The exoskeleton recorded a person walking to a table, after which the robot learned to walk, despite having no similar examples in its training data.

Experts reported that the robot successfully completed all mobility tasks and continued to interact with objects without losing its balance.

In one of the tests, researchers physically pushed the robot, which autonomously returned to its original position and completed the task.

Additionally, it is worth reminding that in August, Nvidia launched a new robotics module, Jetson AGX Thor, priced at $3499, which the company refers to as the «brain of the robot.»