In brief
- LiDAR, sensors, and AI models convert exoskeleton motion into robot-ready actions.
- Study shows robots can gain new skills from exoskeleton data alone, cutting costs.
- A Unitree G1 learned to walk after just five teleoperated demonstrations.
A research team from China’s National University of Defense Technology and appliance maker Midea Group aims to solve one of robotics’ most challenging problems—teaching humanoid robots to move like humans without relying on thousands of costly demonstrations.
To address these issues, the team introduced HumanoidExo in a research paper published last week. The lightweight wearable suit records a person’s full-body motion (arms, torso, and legs) and converts it into structured data for robot learning.
In tests, a Unitree G1 humanoid robot trained on the data learned to perform complex manipulation tasks and even walk after being exposed to only a few examples.
“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 wrote.
Humanoid robots often fail to generalize human motion because their training data comes from video or simulation. HumanoidExo addresses that gap by capturing real joint-space motion.
The suit maps seven human arm joints directly to a robot’s configuration, uses inertial sensors on the wrists, and adds a LiDAR unit on the back to track the wearer’s torso and height.
That motion stream feeds into a dual-layer AI system called HumanoidExo-VLA, a Vision-Language-Action model that interprets the task and a reinforcement-learning controller that maintains balance during movement.
The Unitree G1 was trained with only five teleoperated demonstrations and 195 exoskeleton-recorded sessions, the researchers said. The hybrid data boosted success on a pick-and-place task from 5% to around 80%, nearly matching a 200-demo baseline.
When the exoskeleton captured a person walking to a table, the robot learned to walk, even though its direct training data contained no walking.
The researchers also claim that the robot achieved a 100% success rate in the locomotion portion and could continue manipulating objects without losing balance.
In one test, researchers physically pushed the robot away from its work area. It recovered by walking back to its position and completing the task.
The study arrives amid a global rush in humanoid robot research.
NVIDIA’s Project GR00T, Google DeepMind’s Gemini Robotics, and startups like Figure AI are racing to scale robot training.
Meanwhile, Paris-based exoskeleton maker Wandercraft, which showcased its Atalante X suit at the 2024 Summer Olympics, has also pivoted toward humanoid robots, launching its new humanoid robot, Calvin 40, in June.
The new robot is based on the company’s easier exoskeleton design.
“We’re seeing humanoid robots everywhere—in the U.S., in China, from Tesla, from Figure AI,” Wandercraft CEO Matthieu Masselin previously told Decrypt.
“For us, it’s the same technology we’ve been developing for the last 10 years, he said. “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.”
Still, the HumanoidExo approach suggested a more accessible path to training humanoid robots, one where teaching a robot to walk could soon mean simply putting on a suit and going for a walk.
Generally Intelligent Newsletter
A weekly AI journey narrated by Gen, a generative AI model.
