Animal-style AI robots learn to navigate unfamiliar terrain

Leeds University News | Peer Review | Embargoes until 10am on Friday, July 11th, 2025

Photos and videos included

Researchers have developed an artificial intelligence (AI) system that allows a four-legged robot to adapt its gait to different, unfamiliar terrain, like a real animal.

Pioneering technology allows robots to change the way they move autonomously, rather than needing to tell them when to change their stride, like current generations of robots. This advancement is seen as a major step towards the use of potentially legged robots in dangerous environments where humans can be at risk, such as nuclear decommissioning and search and rescue.

For research conducted by the University of Leeds and the University of London (UCL), researchers took inspiration from the Animal Kingdom and taught robots to navigate terrain that they had never seen before. This included four-legged animals, including dogs, cats and horses. These animals switch ways of moving to conserve energy, maintain balance, and respond quickly to threats.

Researchers have created a framework that can teach robots how to transition, such as running, running, and boundaries, as mammals do in nature.

Switch walking if necessary

By embedding in AI systems, robots quickly learn to switch walking on the spot, depending on the terrain, by embedding the same strategies that animals use to navigate the unpredictable world. Thanks to the data processing power of AI, robots (the robot called “Clarence”) have learned the necessary strategies in just nine hours, much faster than most young animals take to pass through different surfaces with confidence.

A paper published today (July 11th) Nature Machine Intelligence, Joseph Humphreys, a graduate researcher at the Leeds School of Mechanical Engineering, explains that it overcomes a variety of topography, including materials, loose wood chips, and loose vegetation, allowing robots to change strides according to the environment, without changing into the system itself.

He said: “Our findings could have a significant impact on the future of robotic motion control of the legs by reducing many of the previous limitations on adaptability.”

He added: “This deep reinforcement learning framework teaches walking strategies and behaviors inspired by real animals or “bio-wind” such as energy savings, movement adjustments when needed, and walking memory.

“All training is done in simulation. Train your policy on a computer, then take it and place it on the robot. He's as skilled as training. It's similar to the matrix when Neo's martial arts skills are downloaded to the brain, but he doesn't get physical training in the real world.

“We then tested real-world robots on surfaces we had never experienced before and navigated them all well. It was really rewarding to adapt to all the challenges we set up and see the animal behavior we learned became almost a second nature.”

Deep Reinforcement Learning Agents are good at learning specific tasks, but have a hard time adapting to change environments. The animal's brain incorporates structures and information that support learning. Some agents can mimic this kind of learning, but those artificial systems are usually neither sophisticated nor complex. Researchers say they overcame this challenge by incorporating natural animal movement strategies into the system.

They say that it is the first framework to integrate all three key components of animal migration into all three key components into all reinforcement learning systems. This means that walking transition strategies, walking procedural memory, and adaptive motion adjustments can be directly enabled for real deployments directly from the simulation without the need for further adjustment of the physical robot.

Simply put, robots don't just learn how to move. Learn which walks to use, when to switch, and how to adjust in real time.

Professor Zhou, senior author of UCL Computer Science's research, said: “This study was driven by basic questions. If a robot with legged can instinctively move the way an animal is done, instead of training the robot for a particular task, I wanted to provide strategic intelligence to use animals using principles such as balance, coordination, and energy efficiency.

“By embedding these principles into AI systems, we have allowed robots to choose how to move based on real-time conditions rather than pre-programmed rules, meaning they can safely and effectively navigate unfamiliar environments.

“Our long-term vision is to develop embodied AI systems (including humanoid robots) that are fluid, resilient, and mobility, adapted and interacted with animals and humans.”

Real World Applications

Engineers are increasingly mimicking nature known as biomimicry to solve complex mobility challenges. The team says their achievements are showing great progress in enabling leg robots to be more adaptable and handle real-world challenges in dangerous environments or difficult to access. Robots that can navigate unfamiliar and complex terrains open up new possibilities for use in disaster response, planetary exploration, agriculture and infrastructure inspections.

It also suggests a promising pathway for integrating biological intelligence into robotic systems and conducting more ethical investigations of biomechanical hypotheses. Instead of burdening animals with invasive sensors or being at risk to study stability recovery responses, robots can be used instead.

Taking inspiration from factors that make animal movement effective, researchers were able to develop a framework that allows robots to pass through complex, high-risk terrain, despite the absence of appearance receptive sensors.

Parallel practice of multiple terrains

Deep Reinforcement Learning – Effectively Super Powerful Trial and Error – Robots solve the challenge of practicing simultaneously within hundreds of environments, first moving on different walks, then selecting the best walking for the terrain, and generating tools to achieve highly adaptive movements.

To test this acquired adaptability in the real world, the robots tested their ability to recover from travel by repeatedly cleaning their legs, loosening them on real life surfaces such as wood chips, rocks, overgrown roots, and loose wood. The team used programmed routes or joysticks to direct the robots, similar to those used in video games.

Perhaps surprising, the robots were not exposed to rough terrain during training, highlighting the system's adaptive capabilities, demonstrating that these skills are instinctive for the robot.

This research is funded by the Royal Society and the Advanced Research and Inventional Institution (ARIA), and focuses on enabling robust everyday movements. In future work, teams hope to add more dynamic skills, such as long distance jumps, mountain climbing, and navigating steep and vertical terrain.

This framework has so far been tested only on single dog-sized quadruped robots, but the underlying principles apply widely. As long as they share similar forms, the same bio-like metrics can be used on a wide range of four-legged robots, regardless of size or weight.

More information

A paper that learns to adapt through bio-wind walking strategies for versatile quadruped power. Nature Machine Intelligence Friday, July 11th, 2025.

doi:10.1038/s42256-025-01065-z

For media enquiries or interview requests, please contact Deb Newman via d.newman@leeds.ac.uk and copy it at pressoffice@leeds.ac.uk.