Researchers trained the four-legged robot to switch between trotting and bounding based on the obstacles ahead. (Image: Jun-Gill Kang)A team of researchers has developed a new artificial intelligence training method that enables a four-legged robot to adapt its movement in real time, allowing it to climb stairs, navigate forests, and leap over obstacles without instructions from a human operator.The robot, called KAIST HOUND, weighs around 45 kg and uses onboard cameras and lidar sensors to analyse the terrain ahead. Based on what it detects, it can automatically switch between two different gaits: a steady trot for stability and a faster bound for tackling larger obstacles.The research, published in the journal Science Robotics, introduces a training framework known as Action Pretrained Transformer-Based Reinforcement Learning (APT-RL). The system combines transformer-based AI with reinforcement learning, enabling the robot to learn movement patterns from pre-generated examples before refining its behaviour through trial and error. The new AI framework combines transformer models with reinforcement learning to improve robotic locomotion. (Image: Jun-Gill Kang)To train the robot, researchers first created around 180,000 simulated trotting and bounding sequences using trajectory optimisation. Although the dataset represented over 15 hours of movement, it took only about eight minutes to generate.The AI was then trained in simulated environments featuring stairs, hurdles, stepping stones, gaps and uneven ground. Unlike traditional robotic systems that rely on separate controllers for different movements, the new approach allows the robot to transition smoothly between gaits without losing balance.Also Read | VLC creator turns attention to robots, raises $5 million for real-time device platformDuring real-world testing, KAIST HOUND successfully completed a 1.1-km campus route and a 300-m forest trail filled with roots, fallen logs and slippery leaves. In another demonstration, the robot bounded over a 60-cm obstacle, briefly reaching speeds of 15 kmph, and safely descended a three-step staircase.Researchers found the robot generally preferred trotting on uneven terrain at lower speeds but automatically switched to bounding when encountering larger steps, gaps, or obstacles. The dual-gait system consistently outperformed versions of the robot restricted to one style of movement.Story continues below this adAlso Read | Airport of the future? Tokyo tests humanoid robots for ground operationsThe team believes this technology could eventually improve the performance of robots deployed in disaster response, search-and-rescue missions, and other environments difficult for wheeled machines to access.While the system supports only two forward-moving gaits, researchers say future work will focus on enabling sharper turns, sideways movement, and additional locomotion modes such as crawling, further expanding the robot’s ability to operate in complex real-world environments.