Robot Learns Animal-Like Locomotion Using AI and Simulation

A new control technology enables a four-legged robot to autonomously choose its locomotion strategy, similar to how animals adapt their movement. Developed by a team at the Korea Advanced Institute of Science and Technology (KAIST), this system allows a single controller to seamlessly switch between gaits like walking, running, and jumping in real-time, facilitating stable and swift movement across varied outdoor terrains.

Existing four-legged robots often struggle with the unpredictable nature of real-world environments, where obstacles like stairs, gaps, and uneven surfaces combine. Previous robots typically required individual control for each gait, limiting their ability to transition naturally between them. This new approach, named APT-RL (Action Pretrained Transformer-based Reinforcement Learning), aims to overcome these limitations by enabling the robot to learn and freely combine various locomotion skills.

The APT-RL technology was trained using extensive data generated through computer simulations, a process significantly faster than traditional motion capture methods. This simulated data, covering a range of gaits, was used to teach the robot fundamental movement capabilities based on its dynamics and trajectory optimization. Subsequently, reinforcement learning was applied to allow the robot to autonomously select the most appropriate gait for complex three-dimensional terrain.

Equipped with a depth camera and LiDAR, the robot, named KAIST HOUND, can perceive its surroundings and target speed in real-time, informing its gait selection. During testing on both indoor obstacle courses and outdoor campus and forest trails, KAIST HOUND demonstrated stable navigation across urban and natural terrains, including stairs, grass, slopes, fallen trees, and leaf-covered paths. The robot achieved a peak speed of six meters per second, showcasing its ability to maintain both speed and stability.

The experiments confirmed that KAIST HOUND could autonomously switch between gaits like trotting and bounding based on terrain and desired speed. The research team believes this foundational technology will expand the applications of AI-powered robots in challenging environments such as disaster sites, defense missions, and industrial inspections.