Watch: Swiss Researchers Develop AI-Powered Robot That Plays Badminton Against Humans

NDTV

3 days ago

Quick Read
Summary is AI generated, newsroom reviewed.
Researchers at ETH Zurich created an AI-powered robot that plays badminton against humans.
ANYmal-D is a four-legged robot equipped with a dynamic arm and stereo camera for gameplay.
Using reinforcement learning, the robot effectively predicts shuttlecock trajectories and returns shots.
Researchers at a Switzerland university have developed an AI-powered legged robot that plays badminton against humans with impressive agility. Researchers at ETH Zurich tested their AI controller on ANYmal-D, a four-legged robot equipped with a stereo camera and a dynamic arm holding a badminton racket, Independent reported. Using reinforcement learning, the robot accurately tracks the shuttlecock's flight, predicts its trajectory, and navigates the court to intercept and return shots. The robot learns by trial and error to make better decisions.
The main adaptation was giving it four legs instead of the two used by humans. The configuration gives the robot much more stability and flexibility in movements.
Watch a robot play badminton against human players.
Learn more: https://t.co/bTB1qXC8Qp @NewsfromScience pic.twitter.com/yEu6YxsM6C
— Science Magazine (@ScienceMagazine) June 3, 2025
According to researcher Yuntao Ma, this project highlights AI's potential for enabling legged robots to perform complex tasks, potentially leading to advancements in autonomous and intelligent systems, including humanoids.
"We introduced a perception noise model that maps the robot's motion to perception quality. And this allowed the reinforcement learning algorithm to automatically balance between the robot's agile motion and a reliable perception,'' said Mr Ma.
"The control algorithm also generalises to other robot platforms such as humanoids and also other tasks such as search and rescue, and home services,'' he added.
What happened in the test game?
The robot designed for badminton was tested against human players, successfully returning shots at various speeds and angles, achieving rallies of up to 10 hits. Using reinforcement learning, the robot tracked shuttlecocks at speeds of 12.06 m/s, adjusting its gait and balance. The robot was also able to rise on its hind legs for better visibility while prioritising balance.
However, it struggled with fast, aggressive shots like smashes due to hardware limitations in camera perception and actuator speed, with a 0.375-second delay in response. Future improvements in perception responsiveness are needed for competitive full-court play.
"A key advantage of our approach is that the controller is trained end-to-end—upper and lower limbs are optimised together from the beginning. There is no architectural distinction between coordinating the arm with the legs and coordinating the left and right legs. As a result, the limbs learn to compensate for each other's dynamics naturally during training, leading to coordinated whole-body motion", he told Interesting Engineering (IE).