Future sport activity sees 4-legged robot ANYmal play badminton with humans

Scientists at ETH Zürich have successfully trained a four-legged robot called ANYmal to play badminton against human opponents, achieving rallies of up to 10 shots through advanced AI and reinforcement learning. The breakthrough demonstrates how quadruped robots can master complex sports requiring coordinated whole-body movement and real-time visual perception, potentially opening new applications for disaster relief and debris removal operations.

What you should know: ANYmal combines sophisticated hardware with AI-powered learning to compete in dynamic sports scenarios.

• The robot weighs 110 pounds, stands 5 feet 3 inches tall with an attached arm, and operates 18 joints simultaneously across four legs and one arm.
• A stereo camera system enables real-time shuttlecock tracking and trajectory prediction during gameplay.
• Through 50 million simulation trials, the robot learned to coordinate all joints while balancing visual perception with agile movement.

How the training worked: Researchers used reinforcement learning to teach ANYmal badminton skills through trial and error in virtual environments.

• The robot practiced in simulated badminton courts, learning to track shuttlecocks served from various positions and angles.
• A virtual coaching system rewarded proper racket positioning, swing angles, timing accuracy, and efficient court movement.
• The neural network created from these simulations was then transferred to the physical robot for real-world testing.

Performance results: ANYmal adapted its movement strategies based on shuttlecock distance and developed surprisingly sophisticated gameplay tactics.

• The robot achieved swing speeds of approximately 39 feet per second—roughly half the speed of amateur human players.
• At distances under 2 feet, ANYmal remained stationary; at 5 feet, it scrambled using all four legs; at 7 feet, it galloped with periods of elevation extending its reach by 3 feet.
• The robot spontaneously began returning to center court after each hit, mimicking human player positioning strategies.

Current limitations: The research team identified several areas where ANYmal’s performance could improve compared to human players.

• The robot doesn’t consider opponent movements when predicting shuttlecock trajectories, unlike experienced human players.
• “Controlling the robot to look at the shuttlecock is not so trivial,” study co-author Yuntao Ma explained, noting the trade-off between visual tracking and movement speed.
• Adding a neck joint could allow longer shuttlecock monitoring periods during gameplay.

Why this matters: The research advances robotics capabilities beyond sports into practical real-world applications requiring dynamic coordination.

• The technology could support debris removal during disaster relief efforts, where robots need to balance visual perception with agile motion.
• Four-legged robots with coordinated limb control could navigate challenging terrain while manipulating objects in emergency situations.
• “Sports is a good application for this kind of research because you can gradually increase the competitiveness or difficulty,” Ma noted.

What they’re saying: Researchers emphasized the complexity of coordinating multiple independent systems into fluid movement.

• “This trade-off has to happen in a somewhat intelligent way,” Ma said about balancing visual tracking with movement speed.
• Ma expressed surprise at “how well the robot figured out how to move all 18 joints in a coordinated way,” noting the challenge of independent motor learning requiring synchronized execution.