On July 6, inside a laboratory at the Institute of Humanoid Robotics, University of Science and Technology of China (USTC), Cai Haorui squatted beside the test field and gently set a silver-grey quadruped robotic dog on a sloped platform simulating lunar soil. He pressed a button on the remote control, and the robot’s four limbs suddenly folded inward as if stripped of their rigid frames, curling its entire body into an almost perfect cylinder. Seconds later, it rolled silently down the incline. Upon reaching the bottom, it automatically unfurled its legs and stood steady, as if nothing had happened.

“It consumes zero power going downhill—it runs entirely on gravity,” Cai Haorui stood up, brushing dust off his hands and speaking matter-of-factly, as if describing a trivial everyday occurrence.

Yet this single movement took him and his team two full years to perfect.

Cai Haorui is a doctoral student at USTC’s Department of Precision Machinery and Precision Instrumentation. In early 2024, his lab received a research mandate for deep-space exploration: designing an innovative robot for prospecting craters at the Moon’s South Pole.

Permanently shadowed regions at the lunar South Pole hold water ice deposits, a critical resource for future crewed lunar outposts. However, the terrain there is far more treacherous than the Moon’s near side—densely dotted with steep-sloped craters and jagged crater floors. Traditional wheeled rovers such as Curiosity and Perseverance can barely traverse such terrain. While legged robots can cross trenches and uneven ground, their numerous joints result in high energy consumption. On the Moon, solar power supplies are limited, and a complete power loss would terminate a mission outright.

“We wondered back then: could we create a robot capable of traversing rugged terrain while saving energy on flat ground and downhill stretches?” Cai said.

The idea sounded promising, but the industry consensus held that legged and wheeled/rolling robots represent two separate technical pathways, each with distinct tradeoffs. Forcing a hybrid design usually compromises both functionalities. Worse still, mature deformable rolling quadruped robot solutions were scarce domestically and globally at the time, offering little reference for their work.

Undeterred, Cai combed through extensive academic literature and found natural blueprints waiting to be studied.

“Take pangolins—they curl into a tight ball when threatened, and they are quadrupedal creatures,” Cai pulled out a photo of a curled pangolin on his phone. “Then there’s the golden wheel spider, a desert species that transforms itself into a wheel to roll downhill when fleeing, far faster than running.”

He distilled the locomotive behaviors of these two creatures into a core design principle: the robot must walk normally and deform to roll, using an identical set of limbs without adding extra mass.

This insight directly shaped the robot’s mechanical configuration. Its four legs function both as walking actuators and rolling spokes. Precision joint control allows the machine to shift from a four-legged standing posture into a cylindrical form. When rolling, the limbs tuck inward, while the outer shell bears friction and impact from the ground.

“Our technical approach sets us apart from all other designs,” Cai emphasized. Most existing rolling robots require additional drive components, such as internal eccentric wheels or externally mounted tires. By contrast, this robotic dog shares one unified joint drive system for walking, deformation and rolling, eliminating redundant weight and delivering superior energy efficiency.

Lab test data shows the robot’s power draw in quadruped walking mode matches that of conventional four-legged robots. In downhill rolling mode, energy use drops nearly to zero—only minimal power sustains the control system. On level ground during active rolling, its power consumption is merely one-fifth of that required for walking.

“On the Moon, energy equals operational lifespan. Every watt-hour saved extends travel distance by one kilometer and enables more data collection,” Cai explained.

He ran the numbers: if 30% of a prospecting mission’s route consists of downhill slopes or flat, hard terrain, switching to rolling mode can more than double the robot’s overall endurance. That means a rover originally limited to operating for one lunar day could function for two, or cover twice the exploration area.

The robot has passed validation in multiple simulated real-world scenarios. The team has also connected with relevant aerospace research institutions, with plans to deploy the machine as a vanguard surveyor in upcoming lunar exploration missions.

Cai leads a young team of 8 to 10 researchers with an average age under 26, drawn from USTC’s precision machinery, control science, artificial intelligence and other disciplines.

The core team members spent nearly all their weekends and holidays in the lab over the past two years. Through countless failed prototypes, they iterated and rebuilt the design until the first functional robotic dog prototype took shape.

The team is now tackling their last major technical hurdle: stable terrain perception during rolling. As the entire robot body rotates while rolling, its vision and radar sensors experience violent jitter. Capturing steady environmental data is a prerequisite for autonomous navigation and decision-making.

“Once we resolve this challenge, we can roll out this technology for a much wider range of practical applications,” Cai said.