INFOGRAPHIC
Walking Before They Run
Two-legged robots that move like humans have been a long-sought-after goal. Engineers are still grappling with how to make humanoid robots that walk with style.
Tiangong Ultra won the robot division of the Beijing half marathon in April 2025. Photo: Getty
In April 2025, around 12,000 competitors ran in a half-marathon in Beijing. Of that group, 21 were especially noteworthy: They were humanoid robots in what was called the first footrace where humans and machines were pitted against each other. Only six of the 21 finished the 13-mile course, and even then their pace couldn’t match that of an average human runner. According to a report in Wired, the fastest robot, called Tiangong Ultra (pictured above), needed three battery changes and fell over once before finishing the course in two and two-thirds hours.
For more than 100 years, robots have been idealized as mechanical copies of humans—with two arms, two legs, and a head—even if that body plan doesn’t make much sense. And companies have been developing walking robots for years as a means of showing how advanced their technology is. (In more than one case, companies have demonstrated their walking robots by sending out human performers in robot suits.)
It turns out that walking on two legs is an incredibly complicated task.
How People Walk
Walking involves more than putting one foot in front of the other. People instinctively shift their bodies to maintain their center of gravity between their feet, and the vestibular sensory system in the inner ear combines with visual cues to maintain balance. These are important as walking involves standing on one leg briefly as the opposite leg swings forward to catch the body as it falls forward. Anything that disturbs those systems can leave people sprawled on the ground and unable to walk.
Robotic First Steps
In flat, controlled settings, two-legged robots could be programmed to walk—or even perform backflips. But the great advantage of walking versus rolling on wheels is the ability to handle rough and irregular terrain, and in those conditions, the rigid programming of robots was insufficient. Even today, walking robots have difficulty outside of the lab, which is why so many of the entries in the Beijing half marathon walked with unnatural looking gaits ... when they weren’t simply falling over.
Does Hardware Matter?
Ghost Robotics, makers of Q-UGV (pictured), developed a system that used resistance to the leg actuator motors to provide feedback in its four-legged dog-like robots. According to a 2024 paper in Science Robotics, while mechanical systems are physically superior to those found in animals, "If you compare animals to robots at the whole system level, in terms of movement, animals are amazing,” said coauthor Tom Libby, senior research engineer at SRI International, in a press statement. “And robots have yet to catch up.”
Holistic Movements
Roboticists are improving control systems. In recent years, many teams have been leaning on sophisticated machine learning tools to enable their robots to “discover" how to walk. For instance, a group at the University of California pretrained a model to walk across a simulated flat surface, then added unevenness to the simulation; when this control model was installed in an actual machine, the robot was capable of hiking along dirt roads and across fields. Startups such as Figure.AI are following the same path.
Keeping It Simple
Other roboticists are exploring how to improve robotic walking by going in the opposite direction. A team of engineers at Carnegie Mellon University have developed a steerable walking robot that requires only one motor to operate. The team derived the design of the robot, called Mugatu, from LEGO minifigs and rounded the bottom of its feet to enable it to shift its center of balance effortlessly. With wobble, stiff-legged gait, Mugatu isn’t going to win any races, but its simplicity could be a step in the right direction.
© 2025 The American Society of Mechanical Engineers. All rights reserved.