Humanoid robots may soon be able to handle rugged trails as well as people do, thanks to new research from the University of Michigan.
A team there has developed an artificial intelligence (AI) system that trains simulated humanoid robots to hike across uneven, obstacle-filled terrain without needing human operators to guide every step.
The new system is called LEGO-H, and it gives robots the ability to plan ahead, keep their balance, avoid hazards, and adjust their stride as the ground changes.
Instead of just walking in a straight line, the robots can hop, step over rocks, squeeze through narrow spaces, or even recover from a stumble—skills that could be crucial for search and rescue operations, ecological monitoring in remote areas, and other missions where human access is difficult or dangerous.
“Until now, most humanoid robots have been almost blind, relying on humans to make their movement decisions,” said Stella Yu, senior author of the study.
“Our model is the first that lets a humanoid robot see, decide, and move entirely on its own.”
To test LEGO-H, the researchers used simulated robots designed by Unitree Robotics.
These virtual robots came in two sizes—an adult about six feet tall and a smaller, child-sized version.
Each was equipped with cameras for vision, basic GPS directions, and a sense of body awareness.
Instead of following step-by-step instructions, the robots were simply told the direction of a destination, such as “0.3 miles northeast,” and had to figure out the best way to get there.
The virtual robots practiced on five different types of trails, each with multiple levels of difficulty.
Their performance was judged on how safely, efficiently, and completely they could reach the goal.
Surprisingly, when compared with robots that were given perfect maps and full environmental data, the autonomous hikers often did just as well—or even better—especially in terms of safety. Their built-in awareness of their own body movements played a big role in preventing falls and damage.
As the robots trained, they developed natural strategies that were never programmed by humans. For example, they leaned sideways to fit through tight spots, walked around large obstacles, or stepped over smaller ones. Most remarkably, they were able to regain their balance after tripping—something past humanoid robots could not do.
So far, the experiments only trained the robots’ legs. The researchers kept their upper bodies fixed because adding arm movements makes the problem much more complex. But now that the leg-focused system works, the team plans to expand it into full-body hiking to boost stability and safety even further.
The ultimate goal is to take these lessons from virtual trails into the real world, putting humanoid robots on actual rugged terrain.
If successful, robots could soon be hiking beside us—not as companions, but as helpers in some of the toughest environments on Earth.
