Korea Advanced Institute of Science and Technology (KAIST) said on the 19th that a research team at the Humanoid Robot Research Center has developed a humanoid robot that can run at 12 kph and does not lose its balance even with its eyes closed or on bumpy roads. It even performs advanced moves such as duck walk and Moonwalk, and is expected to be deployed at industrial sites in the future.
A research team led by Professor Park Hae-won of KAIST's Department of Mechanical Engineering and Humanoid Robot Research Center (Hubo Lab) developed a lower-body robot platform, 165 cm tall and 75 kg in weight, similar in size to a person, tailored to human-centered environments. They independently designed and produced all core robot components, including motors, reducers, and motor drives, securing key parts in-house.
The researchers trained a robot controller with a reinforcement learning algorithm they developed in-house in a virtual environment and then applied it as is to an experimental robot. In other words, they implemented in the real world what was learned in simulation. The robot can currently run at about 12 kph on flat ground and can overcome a height difference of 30 cm. The team plans to raise performance to 14 kph running, climbing ladders, and overcoming height differences of 40 cm or more.
This humanoid was co-developed by researchers in Korea and abroad. KAIST Professor Hwangbo Jemin took charge of the arms, and Massachusetts Institute of Technology (MIT) Professor Kim Sangbae handled the hands, while KAIST Professor Myung Hyun developed localization and mobility, and Professor Lim Jae-hwan developed vision-based manipulation intelligence. Park and these researchers are working together to develop a complete humanoid with an upper body.
Once completed, the robot will be able to perform complex tasks that require walking and manipulation at the same time, such as carrying heavy objects, operating valves and doorknobs, pushing carts, and climbing ladders, the researchers said. The team aims to equip it with a range of physical abilities sufficient to work alongside people in real industrial settings.
In this process, the researchers also developed a "hopping robot" that jumps and maintains balance on a single leg. This robot demonstrated extreme athletic abilities, such as a 360-degree somersault. With no biological model to reference, the team implemented a reinforcement learning-based artificial intelligence (AI) controller that reduces landing impact while maintaining speed.
Reinforcement learning is a training method that, rather than repeatedly explaining a specific behavior to a dog, rewards the behavior with praise or food when it is performed. It enables AI to learn, through trial and error, how to achieve goals without being instructed at every step.
Professor Park Hae-won said, "This is an important milestone in that we secured everything from core components to the AI controller with our own technology," adding, "We will evolve it into a complete humanoid that includes the upper body and develop it into a next-generation robot that works alongside people." The results will be presented on Oct. 1 at the International Conference on Humanoid Robots (Humanoids 2025).