Chinese researchers have taught a Unitree G1 humanoid robot how to ride a skateboard with a new physics-aware control system called HUSKY.
China Telecom’s Institute of Artificial Intelligence (TeleAI) developed the framework in collaboration with engineers from Shanghai Jia Tong University, The University of Science and Technology of China, ShanghaiTech University, and The University of Hong Kong. HUSKY is somehow short for: Humanoid Skateboarding System via Physics-Aware Whole-Body Control.
The researchers started by teaching the compact humanoid the basic physics of how a skateboard behaves when a person stands on it. They focused on how pushing creates forward motion and how leaning causes the board to turn. They used a small amount of human motion data to simulate a sense of timing and movement that human skateboarders only get from practice.
The system breaks skateboarding into clear phases: pushing, riding, steering, and stepping on and off the board. It mathematically links how much the robot leans to how sharply the skateboard turns so the robot doesn’t guess or overcorrect. The humanoid practiced the behaviors in a simulated world that respects Earth’s laws of physics before the learning was transferred to the hardware via a process called Sim2Real, short for Simulation to Reality.
During real-world testing, the G1 could repeatedly perform the full skateboarding cycle without human assistance. It could operate on different skateboard designs on both indoor and outdoor surfaces, and remained stable even when lightly disturbed. The researchers found that accurate modeling of the skateboard’s physical properties was critical. When the wrong stiffness or damping values were used, the G1 failed to mount the board and over-leaned during turns.
Looking ahead, the HUSKY team plans to add onboard vision since the current system relies mainly on body sensing and does not actively see the skateboard on the ground. Better visual perception would help the robot react to cracks, slopes, obstacle, and changing surfaces that trip catch even experienced skateboarders off guard.
The team has open-sourced the project in hopes that other robotics labs will explore more complex tasks.