China's Beijing Institute for General Artificial Intelligence has revealed a new framework for humanoid robot movement called "OmniXtreme." This framework allows robots to perform advanced athletic moves. These include backflips and martial arts kicks.
With this framework, humanoid robots can do many complex moves. They have a high success rate in real-world uses.
This framework lets one algorithm control many movements. It greatly improves how well robots train for advanced physical skills. This is according to a report by CGTN, seen by "Al-Arabiya Business."
The humanoid robot "Unitree G1" was used to test the framework. The robot showed many skills. It could do breakdancing. It could also do a fast series of backflips.
The robot did five Webster flips in a row. It performed long breakdancing moves. It also did forward rolls and butterfly kicks. Many other moves were shown.
While most moves were for dancing, the robot also did push-ups. It completed several sets of attacks.
Biggest Challenge for Humanoid Robots
Making robots move in dynamic and coordinated ways is very hard. This is a big challenge in the robot industry. Usually, reinforcement learning is used to teach robots complex moves. But, the more complex the moves, the harder they are to control.
Xiang Huang, one of the study authors for the new framework, wrote on X. He said, "Most extreme dynamic moves you see are done with over-trained following policies. It has been a challenge to train one policy that can do many diverse, non-traditional moves with high success."
The "OmniXtreme" policy framework was made to meet this challenge. Unlike older reinforcement learning methods that train one policy from scratch, "OmniXtreme" uses a two-stage learning framework as a general policy.
With this policy, humanoid robots can do very hard moves. This includes backflips one after another. They can keep dynamic balance with great precision. They can even breakdance with fast changes in contact points. This is according to Huang.
This means the robot can quickly change which body parts touch the ground or surfaces during complex moves.
The system has had success rates over 90% in many high-dynamic tasks. This is according to the Beijing Institute for General Artificial Intelligence.