Learning Slope-Adaptive Whole-Body Locomotion for Humanoid Robots in Roofing Construction
This paper presents a task-semantic, scene-grounded framework that combines retargeted human demonstrations with trajectory optimization and reinforcement learning to enable a Unitree G1 humanoid robot to robustly perform complex, slope-adaptive whole-body locomotion and construction tasks, such as roofing, while maintaining precise support and work-clearance constraints.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a robot that can walk, climb, and work just like a human. For years, scientists have taught machines to move by showing them videos of people walking or running. The computer copies the movements, telling the robot's legs and arms to mimic the human body. This works well on flat floors, but it breaks down when the ground tilts. On a steep roof, a simple copy of a human motion can be dangerous. If the robot tries to copy a human bending down to hammer a nail, the computer might place the robot's feet in mid-air or push its hands right through the roof tiles. The robot needs to understand not just how the body moves, but where that body is in relation to the specific surface it is standing on. This is the core challenge of giving robots the ability to work in the messy, uneven world of construction.
A team of researchers at the University of Florida tackled this problem by teaching a humanoid robot how to act like a roofer. They focused on a specific type of robot called the Unitree G1, which stands on two legs and has arms like a person. The goal was to move beyond simple walking and teach the machine to traverse a sloped roof, bend down to work, and use tools like a nail gun or a hammer without falling or crashing into the surface. The researchers realized that simply copying a human's movements was not enough. A human knows instinctively how to adjust their balance on a slant, but a robot following a video script might not realize its foot is hovering over a gap or its hand is about to punch through the roof. To solve this, the team created a new method that combines human motion with a precise digital map of the roof.
The process began by recording a human worker performing roofing tasks on a laboratory ramp. The person wore a virtual reality headset and trackers on their hands and ankles, allowing the researchers to capture the timing and flow of the movements. This data was then transferred to the robot, creating a first draft of the motion. However, this draft was flawed. Because the robot's body is different from a human's, the initial copy placed the robot's feet in the wrong spots and its hands at the wrong heights relative to the roof. To fix this, the researchers used a detailed 3D model of the roof surface. They adjusted the robot's reference path so that its feet were firmly planted on the slope and its hands stayed at a safe, specific distance from the tiles. For example, they ensured the robot's hand stayed exactly 25 centimeters away from the roof when holding a nail gun, and 12 centimeters away when using a hammer. This step ensured the robot's plan was physically possible and safe before it ever tried to move.
Even with a perfect plan, a robot can still make mistakes when it actually moves. The real world is full of small slips and delays that a computer simulation does not always predict. To handle this, the researchers taught the robot to pay attention to these safety rules while it was moving. They used a learning method that rewarded the robot for keeping its hands at the correct distance from the roof and punished it if any part of its body touched the surface when it shouldn't. This training happened in a computer simulation first, where the robot practiced thousands of times on different roof angles, ranging from a gentle 9-degree slope to a steep 25-degree incline. The results showed that the robot learned to adapt its posture to the slope, keeping its balance while bending down to work.
The team then tested these skills on the actual robot in their lab. They set up a real roof platform and let the robot try the tasks it had learned in the simulation. The robot successfully walked up the slope, bent down to hold a nail gun, and swung a hammer. In every test, the robot kept its feet on the roof and its hands at the right distance, avoiding any collisions with the surface. The measurements showed that the robot's movements were accurate, with its body staying within a few centimeters of the intended path. The researchers also compared their method to other approaches. They found that robots trained only to follow a reward signal without a human motion guide ended up in strange, crouched positions that did not look like real roofing work. Similarly, robots controlled directly by a human operator in real-time lost their balance and fell when trying to step onto the slope. The new method, which combined human motion, a digital map of the roof, and safety-aware learning, was the only one that worked reliably.
This study suggests that for robots to work in construction, they need more than just the ability to walk. They need to understand the geometry of the surface they are standing on and maintain a safe relationship with it while they work. The researchers found that simply copying human movements is not enough; the robot must be grounded in the physical reality of the job site. By teaching the robot to respect the roof's shape and keep its tools at the right distance, the team created a foundation for machines that could one day help with dangerous tasks like roofing. While the current tests were done in a controlled lab with a safety harness, the results show that the basic skills for climbing and working on a slope are within reach. The next step will be to see if these robots can handle the rough, uneven, and unpredictable conditions of a real construction site.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.