“`html
ANYmal playing badminton versus a person.(Image credit: © 2025 Yuntao Ma, Robotic Systems Lab, ETH Zurich.)ShareShare by:
- Copy link
- X
Share this article 0Join the conversationFollow usAdd us as a preferred source on GoogleNewsletterSubscribe to our newsletter
8a8-be4e-4284-9d84-327dcc236ecf”>Researchers have taught a quadrupedal robot to engage in badminton against a human rival, and it dashes across the playing surface to engage in rallies of up to 10 volleys.
By merging full-body kinetics with visual awareness, the robot, dubbed “ANYmal,” figured out how to modify its locomotion to intercept the shuttlecock and effectively knock it back over the partition, all due to artificial intelligence (AI).
You may like
-
Watch new humanoid robot pirouette, pose and pull off deft karate moves with eerily lifelike movement
-
Watch: Chinese company’s new humanoid robot moves so smoothly, they had to cut it open to prove a person wasn’t hiding inside
-
New ‘Transformer’ humanoid robot can launch a shapeshifting drone off its back — watch it in action
ANYmal is a quadrupedal, canine-esque robot that tips the scales at 110 lbs (50 kg) and has a height of approximately 1.5 ft (0.5 m). Sporting four limbs enables ANYmal and comparable four-legged robots to navigate difficult landscapes and ascend and descend obstructions.
Prior studies have outfitted these canine-like machines with arms and instructed them to retrieve designated items or operate doors by securing the handle. However, synchronizing limb management and visual understanding in a vibrant milieu poses a persistent hurdle in robotics.
“Sports represents a suitable use case for this line of investigation because the level of competition or complexity can be gradually ratcheted up,” stated study co-author Yuntao Ma, a robotics expert formerly at ETH Zürich and currently with the tech company Light Robotics, in an interview with Live Science.
Teaching a new dog new tricks
For this study, Ma and his collaborators affixed a responsive arm wielding a badminton racket at a 45-degree inclination onto the standard ANYmal robot.
Following the arm’s incorporation, the robot measured 5 feet, 3 inches (1.6 m) in height and had 18 pivot points: three on each of its four limbs, and six on the arm. The researchers conceived a sophisticated integrated structure that governed the arm and leg movements.
The group also incorporated a stereo camera, featuring twin lenses positioned one above the other, slightly right of center on the anterior of the robot’s frame. The paired lenses facilitated the processing of visual data concerning the arriving shuttlecocks in real time and ascertaining their projected course.
You may like
-
Watch new humanoid robot pirouette, pose and pull off deft karate moves with eerily lifelike movement
-
Watch: Chinese company’s new humanoid robot moves so smoothly, they had to cut it open to prove a person wasn’t hiding inside
-
New ‘Transformer’ humanoid robot can launch a shapeshifting drone off its back — watch it in action
The ANYmal configuration the researchers used. The four-legged robot has a long dynamic arm holding a badminton racket at a 45-degree angle and a stereo camera attached to the front.
The robot was then coached to evolve into a badminton competitor through reinforcement learning. Using this kind of machine learning, the robot probed its surroundings and employed experimentation to master the art of spotting and monitoring the shuttlecock, maneuvering toward it, and swinging the racket.
To accomplish this, the researchers initially fashioned a simulated environment featuring a badminton court, with the robot’s virtual counterpart situated centrally. Simulated shuttlecocks were dispatched from the vicinity of the opponent’s court division, and the robot was charged with following its position and predicting its trajectory.
Following this, the researchers established a demanding education program to educate ANYmal on how to hit the shuttlecocks, with a digital tutor incentivizing the robot for a spectrum of attributes, including the racket’s position, the racket head’s angle, and the velocity of the swing. Notably, the swing incentives were based on time to encourage precise and punctual strikes.
Given that the shuttlecock could make landfall anywhere within the court, the robot was also rewarded if it traversed the court efficiently and avoided unnecessary acceleration. ANYmal’s mission was to amplify the degree to which it was compensated throughout all the assessments.
Leveraging 50 million iterations of this simulated instruction, the researchers devised a neural network competent in regulating the motion of all 18 joints to advance toward and strike the shuttlecock.
A fast learner
Post-simulation, the scientists transplanted the neural network into the robot, and ANYmal was subjected to practical trials in the physical realm.
At this point, the robot underwent training to locate and monitor a vivid orange shuttlecock dispensed by an alternative apparatus, empowering the researchers to dictate the shuttlecocks’ velocity, angles, and landing locations. ANYmal was required to dart across the court to make contact with the shuttlecock at a tempo that would rebound it over the barrier and to the core of the court.
The scientists ascertained that, subsequent to extensive tutoring, the robot demonstrated an aptitude for tracking shuttlecocks and precisely rebounding them with swing speeds attaining approximately 39 ft/s (12 m/s) — approximately half the swing velocity of a typical amateur badminton participant, according to the researchers.
Furthermore, ANYmal adapted its kinetic patterns in accordance with the distance it needed to traverse to the shuttlecock and the timeframe it possessed to arrive at it. The robot was absolved from traveling when the shuttlecock was slated to settle merely a handful of feet (half a meter) distant, but at roughly 5 ft (1.5 m), ANYmal hustled to intercept the shuttlecock by mobilizing all four limbs. At around 7 ft (2.2 m) of separation, the robot galloped toward the shuttlecock, engendering a period of elevation that amplified the arm’s reach by 3 ft (1 m) toward the intended destination.
“Directing the robot to gaze upon the shuttleclock is not altogether elementary,” Ma articulated. The robot’s mobility is significantly impeded if it is fixated on the shuttlecock. Conversely, lacking visual observation means it cannot discern the requisite course. “This equilibrium must be attained with a semblance of intellect,” he conveyed.
Ma conveyed astonishment at the robot’s adeptness in orchestrating the simultaneous motion of all 18 joints in a harmonious manner. This endeavor presents particular complexity considering that the motor affiliated with each joint acquires knowledge independently, yet the culminating motion necessitates their collaborative function.
RELATED STORIES
—Scientists reveal new hydrogen-fueled ‘robot horse’ that could one day take you up a mountain
—Chinese ‘robot dog’ that moves like a cat could revolutionize space exploration and asteroid mining
—Watch humanoid robots waltzing seamlessly with humans thanks to AI motion tracking software upgrade
The cohort additionally observed that the robot instinctively commenced repositioning toward the court’s epicenter following each strike, mirroring the preparatory conduct of human athletes awaiting ensuing shuttlecocks.
Nevertheless, the scientists remarked that the robot disregarded the adversary’s maneuvers, a pivotal facet of how human participants forecast shuttlecock flight patterns. Incorporating human posture projections would potentially elevate ANYmal’s prowess, as indicated by the team in the investigation. Alterations could entail the incorporation of a neck joint to afford the robot extended visual surveillance of the shuttlecock, Ma stated.
He anticipates that the ramifications of this investigation will ultimately transcend the sphere of athletics. For instance, it could facilitate the retrieval of debris amid disaster assistance initiatives, affording the robot the aptitude to harmonize dynamic visual discernment with nimble locomotion, he suggested.
Sophie BerdugoSocial Links NavigationStaff writer
Sophie is a staff writer based in the U.K. at Live Science. Her purview encompasses a diverse assortment of subjects, having previously chronicled explorations spanning from bonobo intercommunication to the identification of primordial H2O in the cosmos. Her compendium of works has additionally graced platforms such as New Scientist, The Observer, and BBC Wildlife, and she garnered recognition as a finalist for the Association of British Science Writers’ 2025 “Newcomer of the Year” designation in acknowledgement of her independent labor for New Scientist. Prior to her tenure as a science journalist, she concluded a doctoral course in evolutionary anthropology at Oxford University, dedicating four years to probing the variance in tool proficiency amongst chimpanzees.
Show More Comments
You must confirm your public display name before commenting
Please logout and then login again, you will then be prompted to enter your display name.
LogoutRead more
Watch new humanoid robot pirouette, pose and pull off deft karate moves with eerily lifelike movement
Watch: Chinese company’s new humanoid robot moves so smoothly, they had to cut it open to prove a person wasn’t hiding inside
New ‘Transformer’ humanoid robot can launch a shapeshifting drone off its back — watch it in action
Robots receive major intelligence boost thanks to Google DeepMind’s ‘thinking AI’ — a pair of models that help machines understand the world
Humanoid robots could lift 4,000 times their own weight thanks to breakthrough ‘artificial