Engineers develop robot that judges its surroundings and walks, runs, and jumps like an animal

KAIST researchers have developed a learning-based control system for the four-legged robot KAIST HOUND that allows one controller to choose and change among locomotion skills in real time. Called APT-RL, the method is intended to let the robot adapt its movement to surrounding terrain, combining walking, running, jumping, ledge-clearing, trotting, and bounding rather than controlling each gait separately.

The team produced 15.5 hours of training data for varied gaits through computer simulation in eight minutes. It used robot dynamics and trajectory optimization to establish basic movements, then applied reinforcement learning so the robot could select and transition between skills on complex three-dimensional terrain. A depth camera and LiDAR supply real-time information about the surroundings and target speed, which the controller uses to select a locomotion strategy.

In tests on an indoor obstacle course, KAIST campus areas, and forest trails, the robot traversed stairs, grass, slopes, fallen trees, exposed roots, and leaf-covered paths. It changed gaits as terrain and target speed changed, and it demonstrated transitions between a trot, using alternating diagonal legs, and a bound, which uses front and rear leg pairs together. On rugged terrain with obstacles, it reached a peak instantaneous speed of six meters per second, about 22 kilometers per hour, while remaining stable.

The study appeared in Science Robotics. KAIST professor Hae-Won Park led the work, with Jun-Gill Kang and Jaehyun Park as co-first authors. The researchers expect the technology could broaden uses for physical-AI-based walking robots in rugged settings including disaster sites, defense missions, and industrial facility inspections.