An era has arrived in which a robot decides on its own how to walk. A quadruped robot control technology that instantly changes its gait to match its surroundings — walking on stairs, jumping over gaps, and keeping its balance on forest paths — has been developed in Korea.
According to the science community on the 18th, a research team led by Professor Park Hae-won of the Department of Mechanical Engineering at the Korea Advanced Institute of Science and Technology (KAIST), together with Professor Hong Seung-woo of Korea University and the Agency for Defense Development, has developed a core control technology for quadruped robots that integrates various gaits such as walking, running, and jumping into a single controller, allowing them to be freely selected and switched.
Quadruped robots that move on four legs are more robust on rough terrain than wheeled robots. However, in actual outdoor environments filled with various obstacles such as stair steps, gaps, and fallen trees, technical limitations remained. Existing quadruped robots control each walking technique individually, so they could not naturally respond to changing situations beyond running on flat ground or passing simple obstacles.
To secure both speed and stability, the team developed a new artificial intelligence (AI)-based control technology called APT-RL (Action Pretrained Transformer-based Reinforcement Learning). In this technology, the robot learns various walking techniques such as walking, running, and jumping in advance, then selects the most suitable movement on its own in real environments according to the terrain and movement speed.
The learning process was also greatly shortened. Previously, large amounts of human or animal movements had to be collected through motion capture. But the team used robot dynamics and trajectory optimization techniques to generate 15.5 hours of high-quality walking data in just 8 minutes through computer simulation alone. Through reinforcement learning, the robot then learned on its own the optimal walking method even on complex three-dimensional terrain such as stairs, steps, and gaps.
A depth camera and LiDAR were also combined to recognize the surrounding environment in real time. The robot analyzes the shape and distance of obstacles and the target movement speed, then immediately selects the most suitable walking strategy. The key is that it can naturally switch between walking, running, and jumping without changing a separate mode or receiving human commands.
The team verified the performance by applying the technology to their self-developed quadruped robot "KAIST HOUND." Experiments were conducted not only on indoor obstacle courses but also in actual outdoor environments such as the KAIST campus and forest paths.
As a result, KAIST HOUND moved stably by changing its gait to match the situation, not only in urban environments with stairs, grass, and slopes but also on unstructured natural terrain such as fallen trees, exposed roots, and leaf-covered paths. In particular, on rough terrain filled with obstacles, it recorded a peak speed of 6 meters per second (about 22 km/h), proving both fast mobility and stability at the same time.
In the experiments, the robot was also observed selecting and switching on its own between the "trot," which alternates diagonal legs, and the "bound," a leaping gait that uses the front and hind legs together, depending on the movement environment and target speed. This means it can integrally perform various movements such as walking, running, jumping, and overcoming steps within a single controller.
Professor Park Hae-won of KAIST said, "This is an achievement that shows a quadruped robot can recognize complex indoor and outdoor terrain on its own and select and switch walking strategies to match the situation," adding, "It will be able to greatly expand the range of applications for physical AI-based mobile robots in environments difficult for people to access, such as disaster sites, defense, and industrial facility inspection."
Professor Hong Seung-woo of Korea University, who conducted the joint research, also said, "By combining model-based techniques such as robot dynamics and trajectory optimization with reinforcement learning, we greatly improved learning efficiency," and added, "It will become a core foundational technology for next-generation mobile robots that can be used even on unpredictable rough terrain."
Meanwhile, the results of this research were selected as the cover paper for the July issue of Science Robotics, the world's most authoritative journal in the field of robotics, and were published on the 15th.
