"Quadruped Robot Technology Evolves to Independently Decide When to Walk, Run, or Jump"

A quadruped robot technology has been developed that can perceive and judge its surroundings like a living animal, allowing it to walk, run, and jump. Unlike previous quadruped robot controllers, which often relied on a single fixed gait, this technology enables the robot to flexibly switch between multiple locomotion modes within a single controller. As a result, the robot can adapt in real time to complex obstacles and ever-changing outdoor environments.

The quadruped robot 'KAIST Hound,' equipped with APT-RL, is navigating stairs and obstacles outdoors. KAIST

KAIST announced on July 16 that Professor Haewon Park's research team in the Department of Mechanical Engineering has developed a core control technology for quadruped robots, enabling rapid and stable movement in real-world outdoor environments, including rough terrain.

Quadruped robots, which move on four legs, have an advantage over wheeled robots when operating in difficult terrain. However, when outdoors, obstacles such as stairs, uneven surfaces, stepping stones, gaps, and branches can impede movement at any time. Therefore, simply being able to walk or run quickly is not enough to ensure smooth mobility.

In fact, while previous quadruped robots demonstrated excellent performance in running on flat terrain or overcoming simple obstacles, they encountered limitations in achieving both speed and stability when faced with complex obstacles in outdoor settings.

Above all, because walking, running, and jumping technologies had to be controlled individually, it was difficult for these robots to respond swiftly to environmental changes.

The research team addressed these limitations by developing the "Action Pretrained Transformer-based Reinforcement Learning (APT-RL)" control technology. This technology allows the robot to pre-learn various locomotion skills, such as walking, running, and jumping, and then autonomously combine and switch action types in real-world environments.

APT-RL Controller Technology Operation System Diagram. KAIST

During the development of APT-RL, the research team generated 15.5 hours of training data—encompassing diverse skills—using only computer simulations, rather than recordings of real human or animal movements. This process was completed in just eight minutes.

The generated data was used to teach the quadruped robot basic movement skills. The learning process incorporated a “robot dynamics” model—mathematically expressing the principles of robot movement—and a “trajectory optimization” technique for efficiently calculating movement paths, allowing for faster and more efficient acquisition of locomotion skills compared to previous approaches.

Furthermore, by applying “reinforcement learning,” in which the robot learns optimal actions through repeated trial and error, the system enables the robot to autonomously select and switch locomotion skills to match the situation, even on three-dimensional terrain such as stairs, uneven surfaces, gaps, and stepping stones. The integration of a “depth camera,” which measures the distance to objects and provides 3D information, and “LiDAR,” which maps the environment in three dimensions using lasers, further enhanced the technical completeness of the system.

Through this, the research team explained, the quadruped robot can recognize its environment and target speed in real time and independently select the most appropriate locomotion strategy.

The most notable feature of APT-RL is that it enables the robot to autonomously select and transition among various locomotion skills—such as walking, running, and jumping—within a single controller, allowing real-time adaptation to environmental changes.

(From left) Jaehyun Park, PhD candidate at KAIST; Haewon Park, Professor at KAIST; Seungwoo Hong, Professor at Korea University; Jungil Kang, Researcher (at the time of the research at the Agency for Defense Development). KAIST

The research team validated APT-RL by installing it on the quadruped robot “KAIST HOUND” and evaluating its performance. Experiments took place indoors, as well as in real outdoor environments such as the KAIST campus and forest trails. As a result, KAIST HOUND demonstrated stable behavioral patterns, switching locomotion skills in real time to match the situation not only on urban terrain including stairs, grass, and slopes, but also in unstructured natural environments such as fallen trees, exposed roots, and leaf-covered paths.

In particular, in challenging environments with obstacles, KAIST HOUND reached a top speed of 6 meters per second (22 kilometers per hour), proving that it can secure both high mobility and stability even in real outdoor conditions.

Professor Park stated, “APT-RL is a technology that combines basic locomotion skill learning with reinforcement learning, enabling quadruped robots to independently select and adjust behavior patterns in diverse environments such as stairs, uneven surfaces, gaps, and forest trails. Going forward, the team anticipates this technology will also be used in the field of physical AI-based legged robots for rough terrains—such as disaster sites where human access is difficult, defense missions, and industrial facility inspections.”

The research was jointly authored by Jungil Kang, researcher (at the time affiliated with the Agency for Defense Development), and Jaehyun Park, PhD student at the Department of Mechanical Engineering at KAIST, as co-first authors, and by Professor Park and Seungwoo Hong, professor at Korea University, as co-corresponding authors. The results of the study were published as the cover article in the July issue of the international robotics journal Science Robotics.

This content was produced with the assistance of AI translation services.