Researchers in China and Hong Kong have created a new artificial intelligence (AI)-based learning system that enables humanoid robots to quickly rise from a resting position, regardless of their position or surrounding terrain.
While the research has not yet been peer-reviewed, the team published their results on GitHub on February 12, including a paper posted to the arXiv preprint database and a video showing their framework in action.
The video shows the bipedal humanoid standing up while on its back, sitting against a wall, lying on a couch, and leaning back in a chair. The researchers also tested the humanoid robot's ability to right itself on a variety of surfaces and slopes, including rocky roads, glass slopes, and against a tree.
They even tried to hinder the robot by hitting or kicking it when it tried to get up. In each case, the robot can be seen adapting to its surroundings and successfully getting up.
This amazing ability to fall and get back up was achieved by a system called Humanoid Standing-up Control (HoST). The scientists achieved this using a method called reinforcement learning, a type of machine learning in which an agent (in this case, the HoST framework) tries to solve a problem through trial and error. Basically, the robot performs an action, and if it leads to a positive outcome, it is sent a reward signal, which encourages it to repeat the action in a similar situation in the future.
Rising to the occasion
The team's system was somewhat more complex than this, using four separate reward groups for more targeted feedback, as well as a set of movement constraints, including motion smoothing and speed limits to avoid erratic or jerky movements. Vertical thrust was also applied in the initial stages of training to help guide the learning processes.
The HoST framework was initially trained in simulations using the Isaac Gym simulator, a physics simulation environment developed by Nvidia. Once the framework was sufficiently trained in simulations, it was implemented into the Unitree G1 humanoid robot for experimental testing, the results of which are shown in the video.
“Experimental results with the Unitree G1 humanoid robot demonstrate smooth, stable and reliable movements in a variety of real-world scenarios,” the scientists said in their study. “Looking forward, this work opens the way for integrating stance control into existing humanoid systems with the potential to expand their practical applicability.”
Standing up may seem like a natural process to us humans, but it is something that humanoid robots have tried to recreate in the past, as can be seen in videos of fallen robots unable to return to an upright position. Teaching a robot to walk or run like a human is one thing, but for real-world use, they must be able to handle complex situations such as tripping, falling, and other difficulties.
Ian StokesNavigate Social LinksCo-Author
Ian is a freelance writer on science and technology, and formerly the Tech & Entertainment Editor for Live Science & Space.com. With a degree in biology and a PhD in chemistry, as well as experience working at the Institute of Physics Publishing, Ian is on a globetrotting odyssey across scientific disciplines. He's seeing how long they'll let him keep this profile picture.
You must verify your public display name before commenting.
Please log out and log back in. You will then be prompted to enter a display name.
Log out
Sourse: www.livescience.com
