In a dazzling fusion of sport and science, researchers at ETH Zurich have taught a four-legged robot to play badminton, and it’s not just swinging wildly. This mechanical marvel, known as ANYmal, is now capable of tracking, predicting, and striking a shuttlecock with surprising accuracy, thanks to a groundbreaking approach that seamlessly integrates locomotion, perception, and manipulation.
Robots often struggle with tasks that require synchronized movement between limbs and real-time visual feedback, particularly in fast-paced environments such as sports. Badminton, with its rapid shuttlecock flights and dynamic footwork, is a perfect stress test for robotic agility and coordination.
To tackle this, ETH Zurich’s team developed a unified reinforcement learning–based control policy that teaches the robot to move, see, and swing simultaneously. This whole-body visuomotor strategy allows ANYmal to use every joint and sensor to its advantage, coordinating leg movements with arm swings while tracking the shuttlecock mid-air.
One of the key innovations is a perception noise model that mimics real-world camera imperfections. By training the robot in simulation with realistic visual errors, the researchers ensured that ANYmal’s performance in the lab translates smoothly to the court. This clever trick encourages the robot to develop active perception behaviors, such as adjusting its stance or head angle to obtain a better view of the shuttlecock.
A four-legged robotic system capable of playing soccer on a variety of terrains
The system also includes a shuttlecock prediction model, allowing ANYmal to anticipate where the birdie will land. Combined with constrained reinforcement learning, the robot learns to position itself optimally and execute precise strikes, even against human opponents. In tests across various environments, ANYmal successfully maintained its balance and navigated the court with agility, proving that legged mobile manipulators can handle the chaos of real-world sports.
While ANYmal’s badminton skills are impressive, the implications go far beyond recreation. This research opens doors to robotic assistants in dynamic settings, from disaster zones to industrial inspections, where coordinated movement and perception are critical. And yes, maybe one day, robots will join us in friendly matches at the local gym.
Journal Reference:
- Yuntao Ma, Andrei Cramariuc, Farbod Farshidian, and Marco Hutter. Learning coordinated badminton skills for legged manipulators. Science Robotics. DOI: 10.1126/scirobotics.adu3922
