In short
- Lidar, sensors and AI models convert exoskeletal movement into robot-ready actions.
- Study shows that robots can only get new skills from exoskelet data, which saves the costs.
- A Unitree G1 learned to walk after only five teleopered demonstrations.
A research team of the Chinese National University of Defense Technology and Appliance Maker Midea Group wants to solve one of the most challenging problems of robotics – teaching humanoid robots to move as people without trusting thousands of expensive demonstrations.
To tackle these problems, the team Humanoidexo introduced in a research paper published last week. The lightweight portable suit records the movement of a person of a person (arms, torso and legs) and converts it into structured data for robot leather.
In tests, a unitree G1 humanoid robot trained on the data that have been learned to perform complex manipulation tasks and even run after they have only been exposed to a few examples.
“A significant bottleneck in learning humanoid policy is the acquisition of large -scale, diverse data sets, because collecting reliable data from practice remains both difficult and cost prohibition,” the researchers wrote.
Humanoid robots often fail to generalize the human movement because their training data comes from video or simulation. Humanoidexo treats that gap by capturing a real joint space movement.
The suit brings seven human arm connections directly to the configuration of a robot, uses slowness sensors on the wrists and adds a Lidar unit at the back to follow the torso and height of the carrier.
That flow of movement feeds itself in a dual-layer AI system called Humanoidexo-Vla, a model 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 television demonstrations and 195 Exoskelete-Recorded sessions, the researchers said. The hybrid data stimulated the success on a pick-and-place task from 5% to around 80%, almost match a 200 demo basic line.
When the exoskelete conquered a person who walked to a table, the robot learned to walk, although the direct training data did not run.
The researchers also claim that the robot has reached a 100% success rate in the reproduction part and could continue to manipulate objects without losing the balance.
In one test, researchers physically pushed the robot away from his working area. It recovered by walking back to his position and completing the task.
The study arrives in the midst of a global almost in Humanoid Robot Research.
Project GR00T from NVIDIA, Google DeepMind’s Gemini robotics and startups such as figure AI racing to scale robot training.
Meanwhile, the Paris-based exoskelet maker Wandercraft, which showed his Atalante X-suit at the Olympic Summer Games 2024, also turned to humanoid robotsLaunch his new humanoid robot, Calvin 40, in June.
The new robot is based on that of the company Easier exoskelet design.
“We see humanoid robots everywhere – in the US, in China, from Tesla, from Figure AI,” said Wandercraft CEO Matthieu Masselin earlier Decodeer.
“For us it is the same technology that we have developed in the last 10 years, he said.” Once we started to get more requests and people put us in that market, it was logical to develop, next to our exoskeleton, a free and autonomous humanoid robot that trusts the same technology. “
Nevertheless, the Humanoidexo approach suggested a more accessible way to train humanoid robots, a where teaching a robot could mean quickly, simply put on a pack and take a walk.
Generally intelligent Newsletter
A weekly AI trip told by Gen, a generative AI model.
