At South Korea's KAIST, researchers have taught a four-legged robot to move through the world the way animals do — not by following rigid scripts, but by reading terrain and choosing its own gait in real time. The system, called APT-RL, allows the machine to walk, trot, bound, or jump across stairs, forest trails, and broken ground at speeds reaching 22 kilometers per hour, all without pausing between movement modes. Trained entirely in simulation in under ten minutes, the robot represents a quiet but meaningful shift in how machines might one day inhabit the unpredictable spaces that humans a
KAIST robot masters adaptive movement, reaching 22 km/h across complex terrain
The robot interprets its environment and responds accordingly
Mark
Why does it matter that the robot can switch between gaits on its own? Couldn't engineers just program it to walk on stairs and run on flat ground?
Mimi
The problem is speed and safety. If a robot has to pause and switch between separate control systems, it hesitates. In a disaster zone with unstable ground, that hesitation can cause it to fall or fail its mission. The robot needs to flow from one movement to another the way an animal does—instantly, without thinking.
Mark
So the real breakthrough is that it learned to do this without watching real animals move?
Mimi
Exactly. Traditionally, engineers would use motion-capture systems to record how a dog or horse moves, then teach the robot to copy that. That's expensive and slow. KAIST generated all the training data through simulation in eight minutes. The robot learned the physics of movement without ever seeing a living thing.
Mark
But simulation isn't reality. How does the robot handle something it never trained on?
Mimi
That's where reinforcement learning comes in. After the initial training, the robot explores different actions and gets feedback on what works. It learns by trial and error, which gives it flexibility to handle novel situations. It's not just following a script—it's developing intuition.
Mark
What does 22 kilometers per hour mean for a four-legged robot? Is that fast?
Mimi
It's fast enough to matter. Most robots either move quickly but unstably, or move slowly and safely. This one does both. On uneven terrain, that's a major milestone. It means the robot can actually be useful in real environments, not just controlled labs.
Mark
Where would you actually use this?
Mimi
Anywhere humans can't safely go quickly. A collapsed building after an earthquake. A factory with hazardous chemicals. A forest fire zone. The robot can navigate those spaces without a human operator constantly adjusting its movements, which means fewer people in danger.
Mark
Does this mean we're close to robots that move like animals?
Mimi
We're closer than we were. This robot doesn't think about its movements the way we do. It responds to its environment instinctively, which is closer to how animals actually move. But there's still a long way to go before robots have the full adaptability of a living creature.
El Pulso
- Traditional legged robots freeze at the seams — the dangerous pause between a walk and a climb has long been the Achilles heel of machines sent into disaster zones.
- KAIST's APT-RL eliminates that hesitation entirely, giving a single unified AI the authority to flow between gaits the way a dog shifts from trot to bound without deliberation.
- The entire movement library — 15.5 hours of walking, running, and jumping data — was generated in just eight minutes of computer simulation, bypassing the costly motion-capture process that has slowed robotics research.
- Tested on real campus grounds and forest trails, the KAIST HOUND climbed ledges, crossed gaps, and stepped over roots in sequence, reaching 22 km/h on irregular surfaces without sacrificing stability.
- The technology is now pointed toward search-and-rescue, industrial inspection, and defense — environments where terrain is hostile and human operators cannot afford to babysit every command.
At South Korea's KAIST, researchers have taught a four-legged robot to move through the world the way animals do — not by following rigid scripts, but by reading terrain and choosing its own gait in real time. The system, called APT-RL, allows the machine to walk, trot, bound, or jump across stairs, forest trails, and broken ground at speeds reaching 22 kilometers per hour, all without pausing between movement modes. Trained entirely in simulation in under ten minutes, the robot represents a quiet but meaningful shift in how machines might one day inhabit the unpredictable spaces that humans and animals navigate by instinct.
A four-legged robot at South Korea's KAIST can now cross broken terrain the way animals do — deciding on the fly whether to walk, trot, bound, or jump, without pausing to switch between control systems. Led by mechanical engineer Hae-Won Park, the team built a machine that reaches 22 kilometers per hour while maintaining balance on stairs, slopes, and forest trails. The system, called APT-RL, marks a fundamental shift: instead of following rigid instructions, the robot learns to read its surroundings and respond with appropriate movement.
The problem KAIST set out to solve is deceptively simple to state. Traditional legged robots excel at individual tasks, but switching between them creates dangerous delays — a hesitation that can mean failure in a disaster zone. The team wanted a single unified system that would let the machine flow from one gait to another as naturally as a dog or horse does.
Rather than relying on expensive motion-capture recordings, the researchers generated all training data through computer simulation — 15.5 hours of movement patterns produced in just eight minutes. The robot then refined its skills through reinforcement learning, trial and error guided by feedback, allowing it to handle conditions it had never explicitly trained on. Two sensor systems — a depth camera and LiDAR — work in concert to give the machine a continuous, real-time picture of its surroundings.
Tested on the KAIST HOUND across campus and forest trails, the results were striking. The robot trotted when stability mattered, bounded when speed was needed, climbed ledges, crossed gaps, and stepped over barriers in sequence without stopping to recalculate. Its peak speed of six meters per second was achieved not on flat ground, but on irregular surfaces — something most robots cannot manage without sacrificing one quality for the other.
Professor Park has called this foundational technology for expanding what walking robots can do in rugged environments. Published in Science Robotics, the work points toward a future where machines move with the flexibility we associate with living organisms — and where search-and-rescue teams, industrial inspectors, and defense operators can send robots into dangerous terrain with far less need for constant human oversight.
A four-legged robot at South Korea's KAIST can now move across broken terrain the way animals do—by deciding on the fly whether to walk, trot, bound, or jump, without pausing to switch between different control systems. Researchers led by mechanical engineer Hae-Won Park built a machine that reaches speeds of 22 kilometers per hour while maintaining balance on stairs, slopes, forest trails, and other unpredictable ground. The system, called APT-RL (Action Pretrained Transformer-based Reinforcement Learning), represents a fundamental shift in how robots navigate the real world: instead of following rigid instructions, the machine learns to interpret its surroundings and respond with appropriate movement.
The challenge that KAIST solved is deceptively simple to state but hard to execute. Traditional four-legged robots excel at individual tasks—they can climb, they can jump, they can walk smoothly—but switching between these movements creates dangerous delays. A robot might hesitate when terrain changes from flat ground to stairs, and that hesitation can mean failure in a disaster zone or on an industrial inspection site. The team wanted a single unified system that would let the machine flow from one gait to another as naturally as a dog or a horse does.
Instead of relying on expensive motion-capture recordings of animals or humans, the researchers generated their training data entirely through computer simulation. In just eight minutes, they created 15.5 hours of movement data covering walking, running, and jumping patterns. The system learned how the robot's body responds to different forces and motions using mathematical models and efficient path planning. After building this foundation, the robot then refined its skills through reinforcement learning—trial and error, with feedback guiding it toward better choices. This approach meant the machine could handle conditions it had never explicitly trained on.
To navigate the world, the robot relies on two types of sensors working in concert. A depth camera builds a three-dimensional map of nearby objects, measuring distances with precision. LiDAR sensors use laser pulses to detect shapes and obstacles over longer distances. Together, they give the machine a real-time understanding of what surrounds it, allowing it to adjust its movement continuously as it moves.
When tested on the KAIST HOUND robot across campus grounds and forest trails, the system proved itself capable of remarkable fluidity. The machine moved across grass and stairs, navigated uneven terrain with fallen branches and exposed roots, and made its own decisions about which gait to use. It trotted when stability mattered, bounding when speed was needed. It climbed a raised ledge, crossed gaps, and stepped over barriers in sequence without stopping to recalculate. The peak speed of six meters per second—about 22 kilometers per hour—is notable because the robot achieved it while maintaining control on irregular surfaces, something most robots cannot do without sacrificing one or the other.
What makes this work significant beyond the lab is the range of real-world applications it opens. Search-and-rescue teams could deploy machines that navigate debris and unstable ground without human operators constantly adjusting commands. Industrial inspectors could send robots into hard-to-reach areas of factories or infrastructure. Military and defense operations could use machines that move reliably in unpredictable terrain. The technology also reduces the need for constant human oversight and preprogrammed instructions, which means robots can operate more independently in dangerous situations—reducing risk for human responders in disaster zones or hazardous environments.
Professor Park has described this as foundational technology that will expand what walking robots can do in rugged environments. The research, published in Science Robotics, suggests a future where machines move with the flexibility and awareness we see in living organisms. That future is not yet here, but the KAIST HOUND has taken a significant step toward it.
Citas Notables
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— Professor Hae-Won Park, KAIST Department of Mechanical Engineering
Quieres la nota completa?
Lee el original en The Brighter Side of News ↗
Technology