HumanoidExo Turns Human Motion Into Data That Teaches Robots to Walk

A analysis staff from China’s Nationwide College of Protection Expertise and equipment maker Midea Group goals to unravel one among robotics’ most difficult issues—instructing humanoid robots to maneuver like people with out counting on 1000’s of pricey demonstrations.

To deal with these points, the staff launched HumanoidExo in a analysis paper printed final week. The light-weight wearable go well with information an individual’s full-body movement (arms, torso, and legs) and converts it into structured information for robotic studying.

In exams, a Unitree G1 humanoid robotic skilled on the info discovered to carry out complicated manipulation duties and even stroll after being uncovered to just a few examples.

“A major bottleneck in humanoid coverage studying is the acquisition of large-scale, numerous datasets, as amassing dependable real-world information stays each tough and cost-prohibitive,” the researchers wrote.



Humanoid robots usually fail to generalize human movement as a result of their coaching information comes from video or simulation. HumanoidExo addresses that hole by capturing actual joint-space movement.

The go well with maps seven human arm joints on to a robotic’s configuration, makes use of inertial sensors on the wrists, and provides a LiDAR unit on the again to trace the wearer’s torso and top.

That movement stream feeds right into a dual-layer AI system known as HumanoidExo-VLA, a Imaginative and prescient-Language-Motion mannequin that interprets the duty and a reinforcement-learning controller that maintains steadiness throughout motion.

The Unitree G1 was skilled with solely 5 teleoperated demonstrations and 195 exoskeleton-recorded periods, the researchers mentioned. The hybrid information boosted success on a pick-and-place job from 5% to round 80%, almost matching a 200-demo baseline.

When the exoskeleton captured an individual strolling to a desk, the robotic discovered to stroll, despite the fact that its direct coaching information contained no strolling.

The researchers additionally declare that the robotic achieved a 100% success price within the locomotion portion and will proceed manipulating objects with out dropping steadiness.

In a single take a look at, researchers bodily pushed the robotic away from its work space. It recovered by strolling again to its place and finishing the duty.

The examine arrives amid a worldwide rush in humanoid robotic analysis.

NVIDIA’s Challenge GR00T, Google DeepMind’s Gemini Robotics, and startups like Determine AI are racing to scale robotic coaching.

In the meantime, Paris-based exoskeleton maker Wandercraft, which showcased its Atalante X go well with on the 2024 Summer time Olympics, has additionally pivoted towards humanoid robots, launching its new humanoid robotic, Calvin 40, in June.

The brand new robotic is predicated on the corporate’s simpler exoskeleton design.

“We’re seeing humanoid robots in all places—within the U.S., in China, from Tesla, from Determine AI,” Wandercraft CEO Matthieu Masselin beforehand instructed Decrypt.

“For us, it’s the identical know-how we’ve been growing for the final 10 years, he mentioned. “As soon as we started getting extra requests and folks pulled us into that market, it made sense to develop, alongside our exoskeleton, a free and autonomous humanoid robotic that depends on the identical know-how.”

Nonetheless, the HumanoidExo strategy recommended a extra accessible path to coaching humanoid robots, one the place instructing a robotic to stroll may quickly imply merely placing on a go well with and going for a stroll.