The Robot That Decides How to Walk: KAIST's Four-Legged Marvel

Four-legged robots have been the subject of intense research and public fascination. They have trotted across factory floors, navigated simple obstacles, and even performed backflips in carefully staged demonstrations. But despite these impressive feats, they have consistently struggled with the messy, unpredictable reality of the world outside the laboratory. Stairs, gaps, uneven terrain, and sudden obstacles have remained significant challenges, limiting their usefulness in real-world applications.

Now, a research team led by Professor Hae-Won Park from the Korea Advanced Institute of Science and Technology (KAIST) Department of Mechanical Engineering has developed a breakthrough control technology that may finally unlock the full potential of four-legged robots. The new system, called APT-RL (Action Pretrained Transformer-based Reinforcement Learning), allows a single controller to select and switch in real time between walking, running, jumping, and other locomotion skills, enabling the robot to move quickly and stably even in challenging outdoor environments. The research was published in the journal Science Robotics.

The Problem with Traditional Robots

Four-legged robots have long been considered superior to wheeled robots for navigating rough terrain. Their ability to step over obstacles, climb stairs, and balance on uneven surfaces gives them a clear advantage in environments where wheels would struggle. But in real outdoor settings, obstacles such as stairs, ledges, stepping stones, gaps, and tree branches appear in quick succession and in unpredictable combinations. The ability to simply walk and run fast is not enough.

Existing four-legged robots have excelled at running quickly across flat ground or clearing simple obstacles, but they have struggled to maintain both speed and stability in environments where obstacles combine in complex ways. Because walking, running, jumping, and other gaits had to be controlled individually, the robots were also limited in how naturally they could switch between them as conditions changed. This fragmentation of control systems meant that robots could not adapt seamlessly to changing terrain, often slowing down or becoming unstable when faced with unexpected challenges.

A New Approach: Learning from Simulation

To overcome these limitations, the research team developed APT-RL, a control technology designed to enable a robot to first learn a range of locomotion skills such as walking, running, and jumping, and then freely combine and transition among them in real-world environments as the situation demands.

Rather than filming the movements of real people or animals, the team generated 15.5 hours of training data covering a variety of gaits using computer simulations alone, completing the process in just eight minutes. That data was used to teach the robot basic movement capabilities, drawing on robot dynamics, a mathematical model of how a robot moves and trajectory optimization, a technique for calculating the efficient path of movement. This approach is far faster and more efficient than earlier methods that relied on motion capture, a technology that records human or animal movement using sensors.

The team then applied reinforcement learning, an artificial intelligence technique in which an agent learns optimal behavior through repeated trial and error, so the robot could autonomously select and switch gaits suited to complex three-dimensional terrain such as stairs, ledges, and gaps. Finally, the team combined a depth camera, which measures the distance to objects to obtain three-dimensional information, with LiDAR, a sensor that uses lasers to measure the distance and shape of the surrounding environment in three dimensions, enabling the robot to recognize its surroundings and target speed in real time and choose the most appropriate walking strategy.

Putting the Technology to the Test

The team tested the control technology on its own four-legged robot, "KAIST HOUND." The experiments were conducted not only on an indoor obstacle course but also in real outdoor environments, including KAIST's campus and forest trails. KAIST HOUND moved stably across urban terrain that included stairs, grass, and slopes, as well as irregular natural terrain such as fallen trees, exposed roots and paths covered in fallen leaves, switching gaits in real time to match the conditions.

In rugged terrain with obstacles, the robot reached a peak instantaneous speed of six meters per second (approximately 22 kilometers per hour), demonstrating that it can achieve both fast movement and stability in real outdoor environments. The experiments showed that KAIST HOUND autonomously selected and switched between a trot, which alternates diagonal legs, and a bound, a leaping gait using the front and back leg pairs together, depending on the terrain and target speed. Critically, it could integrate walking, running, jumping and ledge-clearing into a single controller, a feat that has eluded roboticists for years.

The Future of Physical AI

Professor Park expressed his excitement about the implications of this research, stating that the team expects this to become a foundational technology that expands the potential uses of physical-AI-based walking robots in rugged environments such as disaster sites, defense missions, and industrial facility inspections.

The research represents a significant leap forward in the field of robotics, demonstrating that robots can now adapt to complex environments with a degree of autonomy that was previously unattainable. By combining advanced simulation training, reinforcement learning, and sophisticated sensor technology, the KAIST team has created a robot that truly understands how to walk on its own.

The study was co-authored by Jun-Gill Kang (affiliated with the Agency for Defense Development at the time of the research) and Jaehyun Park, a Ph.D. candidate in KAIST's Department of Mechanical Engineering, as co-first authors, with Professor Seungwoo Hong from Korea University and Professor Park serving as co-corresponding authors. The research was published in Science Robotics, one of the most prestigious journals in the field, underscoring the significance of this achievement for the future of robotics and artificial intelligence.