Meet 'KAIST Hound,' the robot dog that outruns humans and clears obstacles at 22 kph

A robot dog that sprints through rugged forest terrain with the agility of a wild animal has made its debut.

KAIST announced Thursday that a research team led by Park Hae-won, a professor in the Department of Mechanical Engineering, has developed a core locomotion control technology for quadruped robots — one that uses a single controller to select and switch in real time among walking, running and jumping, enabling fast and stable movement even in outdoor environments.

Existing quadruped robots excelled at running quickly on flat ground or clearing simple obstacles, but struggled to achieve both speed and stability in real-world environments with mixed, complex terrain. They also required separate controllers for each locomotion mode — walking, running and jumping — which made seamless transitions in response to changing conditions difficult.

To overcome these limitations, the research team developed a new learning-based control technology called APT-RL, which stands for action pre-training transformer reinforcement learning.

APT-RL is designed so that a robot first learns a range of locomotion skills — walking, running and jumping — and can then freely combine and switch among them in real-world situations as conditions demand.

Without capturing motion from real people or animals, the team generated 15.5 hours of locomotion training data through computer simulation alone, completing the process in just eight minutes.

The team then applied reinforcement learning so the robot could independently select and switch locomotion strategies on complex three-dimensional terrain, including stairs, steps, gaps and stepping stones. They also integrated a depth camera with lidar, allowing the robot to perceive its surroundings and target speed in real time and choose the most appropriate locomotion strategy.

The research team validated the technology by deploying it on their own quadruped robot, KAIST Hound.

Hound navigated both urban terrain — including stairs, grass and slopes — and unstructured natural terrain such as fallen trees, exposed roots and leaf-covered paths, switching locomotion modes in real time as conditions changed. In rough terrain with obstacles, it reached a peak speed of 6 meters per second, or about 22 kilometers per hour, demonstrating that high mobility and stability can be achieved simultaneously in actual outdoor environments.

Hound autonomously selected and switched between a trot — in which diagonal pairs of legs alternate — and a bound — a leaping gait in which the front and rear legs each move together — depending on the terrain and target speed. The tests confirmed that walking, running, jumping and step-climbing can all be performed within a single integrated controller.

"We expect this to serve as a foundational technology that expands the potential of physical AI-based walking robots in challenging environments, including disaster sites, defense missions and industrial facility inspections," Park said. "Applying this to real missions will require technology capable of autonomous movement over longer distances," he added.

The research was selected as the cover paper of the July issue of Science Robotics, an international academic journal in the field of robotics.