A new AI-powered robotic system inspired by animals is making strides — literally — toward real-world agility, thanks to researchers at the University of Leeds and University College London (UCL).
In what is believed to be a world-first breakthrough, scientists have developed an artificial intelligence framework that allows a four-legged robot to autonomously adjust its gait to tackle unfamiliar and uneven terrain, much like animals in the wild. Their paper was published Friday in the journal Nature Machine Intelligence.
Nicknamed “Clarence,” the dog-sized robot was trained using deep reinforcement learning to transition between different gaits — trotting, running, bounding — based on real-time ground conditions. Unlike conventional robots that require pre-programmed commands to change movement patterns, Clarence autonomously decides how best to walk, run or adapt without any exteroceptive sensors for sight, smell or hearing.
“Our findings could have a significant impact on the future of legged robot motion control by reducing many of the previous limitations around adaptability,” said lead author Joseph Humphreys, a postgraduate researcher at the University of Leeds. “This deep reinforcement learning framework teaches gait strategies and behavior inspired by real animals — or bio-inspired — such as saving energy, adjusting movements as needed, and gait memory, to achieve highly adaptable and optimal movement, even in environments never previously encountered.”
The robot learned its repertoire of motion strategies in just nine hours of computer simulation training — substantially faster than real animals, which can take days or weeks to master movement on different terrain. Once deployed into the real world, Clarence successfully navigated surfaces it had never “seen” before, including loose wood chips, overgrown roots, and uneven timber, and was even able to recover when deliberately tripped.
“It was really rewarding to watch it adapt to all the challenges we set and seeing how the animal behavior we had studied had become almost second nature for it,” Humphreys said.
The study represents the first time a robotic AI system has integrated three key aspects of natural locomotion: procedural gait memory, motion adjustment, and terrain-based gait switching.
“This research was driven by a fundamental question: what if legged robots could move instinctively the way animals do?” said Professor Chengxu Zhou of UCL Computer Science, the studys senior author. “By embedding those principles into an AI system, weve enabled robots to choose how to move based on real-time conditions, not pre-programmed rules. That means they can navigate unfamiliar environments safely and effectively, even those that they havent encountered before.”
Beyond the science, the implications are significant. Legged robots capable of adapting their movements could be deployed in dangerous or hard-to-reach places such as disaster zones, nuclear sites, or even other planets. While legged robots are already deployed by first responders in certain reconnaissance or search-and-rescue operations, they are typically “driven” by human operators and cannot autonomously adapt to unfamiliar terrain.
The researchers say their method also provides a more ethical way to study biomechanics, by replacing animal testing with robotic models that mimic biological movement.
Future work will explore even more complex skills, such as climbing and vertical navigation. While the current framework was tested on a single robot, the underlying principles are transferable across other legged robots of similar design.
“Our long-term vision is to develop embodied AI systems — including humanoid robots — that move, adapt, and interact with the same fluidity and resilience as animals and humans,” Zhou said.