[Wafer: response was truncated before the model finished its internal reasoning. Increase max_tokens, or disable thinking on this model (e.g. chat_template_kwargs.enable_thinking=false), then retry.]

Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have built a four-legged robot that decides on its own whether to walk, run, or jump as terrain changes, eliminating the need to switch between separate control programs. In findings published Aug. 8, the team reported the robot, called KAIST HOUND, reached speeds of up to 6 meters per second, about 22 kilometers per hour, on difficult terrain.

The system, called Action Pretrained Transformer-based Reinforcement Learning, or APT-RL, was developed by a team led by Professor Hae-Won Park of the Department of Mechanical Engineering. The work appears in the journal Science Robotics.

Park said the technology could widen the range of environments where walking robots are used. "We expect 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," Park said.

One system for multiple gaits

Four-legged robots can step over obstacles and climb uneven ground, but real environments present a constant mix of challenges. A robot may face stairs, loose soil, gaps, and fallen debris within seconds, and each obstacle demands a different type of movement. Traditional systems control each gait separately, so robots hesitate when switching between motions. In fast-changing terrain, that hesitation can lead to failure.

APT-RL treats walking, running, and jumping as parts of a single framework rather than separate tasks. The robot shifts between these motions naturally, without manual switching, and reacts to its surroundings in real time.

Training in simulation

Teaching robots to move usually requires large amounts of real-world data, often captured with motion-capture systems that record animal or human movement. The KAIST team generated training data entirely through computer simulation instead. In eight minutes, the system produced 15.5 hours of movement data covering walking, running, jumping, and the robot's physical responses to forces and motion.

After building that base knowledge, the robot improved through reinforcement learning, a trial-and-error method in which it explores actions and receives feedback on its success. This lets it handle conditions that were not part of its initial training.

Sensing and real-world tests

The robot uses a depth camera to measure distances and build a three-dimensional view of nearby objects, plus LiDAR sensors that scan the environment with laser pulses over longer distances. Together, these systems let it map its surroundings and adjust movement in real time.

In tests on campus grounds and forest trails, KAIST HOUND moved across grass, stairs, slopes, fallen branches, and exposed roots. It switched between trotting, which uses alternating diagonal legs for stability, and bounding, which uses paired legs for faster movement. The robot made those choices on its own, without human input.

It also handled sequences of obstacles without stopping, moving from flat ground to a raised ledge to uneven terrain while adjusting its gait smoothly. Most robots sacrifice stability for speed or move slowly to stay steady; this system managed both.

Why this matters for science and research professionals

The APT-RL pipeline offers a template for training legged robots without expensive motion-capture data, an approach that could speed up work in robotics labs and extend to humanoid systems. For researchers studying movement and control, it shows how a single learning system can replace multiple hand-coded controllers. The full findings are available in Science Robotics.