Researchers in South Korea have developed a robot that can walk and run on any type of terrain.
The quadruped robot, RaiBo, can equally keep its balance on grass, a sandy beach, an athletics track or an air mattress.
Its developers say previous robotic systems could walk on shifting terrain, but only walk.
“Walking and running are very different. If you walk very gently, the sand doesn’t deform, and stays intact, which makes it very easy to walk on,” said Professor Jemin Hwangbo, of the Robotics & Artificial Intelligence Lab at the Korea Advanced Institute of Science & Technology (KAIST).
“But when the robot runs, it pushes the ground hard. And then the ground deforms. So it has to account for the deformation,” Hwangbo told Euronews Next.
RaiBo has demonstrated it can run at more than 10km/h and swiftly change direction, turning around at a speed of 90°/s on top of an air mattress.
A new patrol robot to help lifeguards?
RaiBo’s battery can last for about three hours – compared to 1.5 hours for Boston Dynamics’ own quadruped robot, ‘Spot.
However, unlike Spot, RaiBo is just a prototype and not for sale at this stage.
But its developers believe the technology behind it could have a wide range of applications.
Hwangbo said RaiBo would make a great lifeguard patrolling onshore.
“We need more lifeguards on the beach in Korea or around the world. And it’s just too expensive to have people everywhere. We could have robots walk around the shore and alert [us] if there is an emergency,” said Hwangbo.
The robot currently has four legs and no arms. The research team plans to add an arm to it so it could clean streets, deliver goods or assist workers in factories.
“Picking, sorting and trashing in a bin is not so difficult,” Hwangbo said, adding RaiBo should be able to handle such tasks “in the very near future”.
Getting AI to learn by trial and error
RaiBo uses a machine learning technique called reinforcement learning – a method where the robot interacts with the environment to collect data on the results of its actions in any situation and use it to improve how it performs a task.
Hwangbo compared it to how animals and toddlers learn by trial and error.
“For example, babies can’t walk at first and keep losing balance, but they eventually learn to walk, right? For that, babies have to interact with the environment for many years and collect the data,” he explained.
“The robot collects the data from the sand as it walks and processes it through our artificial neural network that predicts what action you should take at that moment”.
As reinforcement learning requires a large amount of data, it often means a long processing time.
However, the research team said it devised a “computationally efficient model” where the robot relies on simulation data instead of real data to decide on its actions more quickly.
RaBbo was featured in the January issue of Science Robotics.
For more on this story, watch the video in the media player above.