A domestic research team has developed a quadruped robot that can freely switch its gait to match the terrain. On flat ground, it moves quickly in a trot by moving diagonally opposite legs at the same time, then when it encounters an obstacle, it jumps over it in a bound by moving the front and hind legs together, respectively. If its ability to find a destination advances, four-legged robots are expected to autonomously carry out search-and-rescue and exploration missions in harsh disaster and combat sites in the future.
A research team led by Park Hae-won, a professor in the KAIST Department of Mechanical Engineering, said on the 16th that it developed an artificial intelligence (AI) control technology for a quadruped robot that can select and switch among various locomotion skills in real time—such as walking, running, and jumping—with a single controller. The results were published the same day as the cover paper in the international robotics journal Science Robotics.
◇ AI brain learns gait with virtual data
A quadruped robot moves with four legs and is theoretically better suited than wheeled robots for traversing rough terrain. But in real environments, the terrain constantly changes and obstacles abound, so it could not move as fast as expected. Because walking, running, and jumping were each controlled separately, it could not smoothly change gaits while running like an actual four-legged animal.
To overcome these limits, the team developed a control technology called APT-RL (Action Pretrained Transformer-based Reinforcement Learning). Without filming every movement of four-legged animals, they created 15.5 hours of data through computer simulation for the quadruped robot's AI brain to perform Reinforcement Learning.
Reinforcement learning is a training method that, rather than repeatedly teaching a dog a specific behavior, rewards it with praise or treats when it happens to perform that behavior. It is widely used in TV entertainment programs when correcting a companion dog's behavior. Through trial and error, the robot independently learned a wide range of locomotion skills much faster and more efficiently than conventional methods and selected among them according to the situation.
The team validated performance by loading the newly developed gait control technology onto its in-house quadruped robot, KAIST Hound. Experiments were conducted not only on indoor obstacle courses but also in outdoor environments such as the university campus and forest trails. The robot moved stably not only in urban terrain with stairs, grass, and ramps, but also in irregular natural terrain where trees had fallen, roots were exposed, or leaves had piled up. In rough terrain with obstacles, it recorded an instantaneous top speed of 6 meters per second (about 22 kph).
◇ Freely switching gaits to match the terrain
The reason the robot traversed rough terrain so quickly was that it could rapidly switch its gait to suit the situation. On flat ground, the robot moved quickly with a trot, alternating diagonally opposite front and hind legs, and when it encountered obstacles, it leaped over them with a bound, using the front and hind legs together, respectively.
The team demonstrated that various motion skills—such as walking, running, jumping, and overcoming height differences—can be performed integratively within a single controller. Cameras and sensors were used together for terrain perception. The robot recognized nearby terrain with a depth camera and checked distant terrain with a lidar sensor, a laser range finder.
Park said, "This study shows that a quadruped robot can recognize complex and unstructured terrain both indoors and outdoors and can independently select and switch to a gait strategy suited to the situation," adding, "Going forward, we expect it to serve as a foundational technology that broadens the potential use of physical AI-based walking robots in harsh environments such as disaster sites, defense missions, and industrial facility inspections."
In this paper, Kang Jun-gil, a researcher at the Agency for Defense Development (ADD), and Park Jae-hyun, a doctoral candidate in the KAIST Department of Mechanical Engineering, are co–first authors, and Park and Hong Seung-woo of the Korea University Department of Mechanical Engineering are co–corresponding authors. The team plans to develop technology that allows quadruped robots to autonomously travel long distances so they can be used in real missions.
Park noted, "If this time we focused on demonstrating fast and diverse locomotion skills, next we will enable the robot to set a destination and plan its own route," adding, "It is also important to develop technology that allows the robot to understand not only obstacles with cameras but also surrounding situations such as people and hazardous objects."
References
Science Robotics (2026), DOI: https://doi.org/10.1126/scirobotics.adz7397