Researchers have created an artificial intelligence method that lets a four-legged robot adjust its motion in real time. The system allows the machine to climb stairs, move through forests, and jump over barriers without commands from a human operator.
The robot, known as KAIST HOUND, weighs about 45 kg. It uses onboard cameras and lidar sensors to scan the area ahead. Depending on the detected terrain, it automatically changes between a stable trot and a quicker bounding gait for larger obstacles.
Published in Science Robotics, the study presents a training approach called Action Pretrained Transformer-Based Reinforcement Learning. It merges transformer models with reinforcement learning so the robot first studies movement examples and then improves performance through repeated practice.
