Humanoid robot plays tennis with real people

“For the first time, a humanoid robot can maintain high-tempo tennis shots for long periods of time with millisecond reflexes, precise ball striking and natural full-body movements,” Chinese company Galbot Robotics wrote on X this week, with a video of Unitree Robotics’ G1 model playing tennis.

The robot achieves this ability thanks to being equipped with Latent, an algorithm jointly developed by Galbot and researchers from Tsinghua University and Peking University. Galbot said Latent is the world’s first real-time full-body planning and control algorithm for a tennis-playing humanoid robot.

Theo Interesting Engineeringone of the main challenges when training sports robots is the lack of accurate data on human movements. This is especially true of tennis, when players have to move over a large area, the ball can reach speeds of up to 30 meters per second, the contact time between racket and ball lasts only a few milliseconds.

To solve the problem, the team avoided recording the entire match. Instead, they focus on capturing short segments of essential movements like forehands, backhands and lateral moves. Data was recorded using a motion tracking system in a 3×5 m soccer field, 17 times smaller than a standard tennis court. A total of 5 players contributed 5 hours of motion data.

Latent first trains the robot to copy individual movements, which are then combined into chains that allow the robot to perform specific tasks including reaching the ball, making a shot and returning to position on the field. To improve real-world performance, the system is trained in a simulated environment, where important physical parameters, such as the mass of the robot and ball, friction and aerodynamics, change randomly.

In the simulation test, the system achieved a 96% success rate with forehand shots. When deployed in practice, the G1 robot demonstrated its ability to maintain back-and-forth ball movements with real people and continuously return the ball to the opponent’s field.

The team notes that this method could be expanded beyond tennis to sports such as soccer, badminton, or other tasks that require sports-related skills but have difficulty collecting full human movement data.

According to Liu Dingding, an Internet analyst in Beijing, the progress shows that humanoid robots have potential applications in households and many fields such as industry and agriculture thanks to their ability to handle complex tasks in each specific context. “This is a major breakthrough for China’s robotics industry as well as its production capacity and technology,” he told Global Times.

The ripple effect could be significant, Liu added, as advances in dynamic control and learning from imperfect data could help reduce training costs, increase scalability, and spur applications in industrial, automation and service environments.