This breakthrough demonstrates that four-legged equipped with can be utilized as competitors in “complex and dynamic sports scenarios.”
How the Robot Dog Became a Badminton Player
The dog-like robot ANYmal, weighing 110 pounds and standing half a meter tall, confidently maneuvers through challenging terrain on its four legs, overcoming obstacles with ease.
Previously, researchers had trained such robots to fetch objects and open doors. However, coordinating limb movement and visual perception in a dynamic environment remained a significant challenge in robotics.
“Sports provide a great domain for this kind of research because you can gradually increase the competitiveness or complexity,” said lead author Yuntao Ma in an interview with Live Science.
His team equipped the robot dog with a dynamic that holds a badminton racket at a 45-degree angle. With the attached arm, the robot’s height increased to 5.2 feet.
Overall, the novice athlete had 18 joints in its limbs. Scientists developed a complex embedded system to control the leg movements. The team also added a stereo camera with two lenses to help the robot process visual information about the approaching shuttlecock and determine its trajectory.
The robot then learned to play badminton through reinforcement learning. Using this type of machine learning, the explored its environment and, through trial and error, detected and tracked the shuttlecock, moving toward it and swinging the racket.
Researchers created a simulation of a badminton court, with a virtual robot at the center. Virtual shuttlecocks were served from the opponent’s side of the court, and the robot had to track and assess their flight trajectory.
Training Stages
Next, researchers developed a rigorous training program to teach the robot dog to return the shuttlecocks. A virtual coach “rewarded” ANYmal for the correct racket angle, swing speed, shot accuracy, and effective movement across the court.
After 50 million attempts, researchers created a neural network capable of controlling the movement of all 18 joints in response to the shuttlecock.
A Step Toward Real Play
Following the simulations, scientists tested ANYmal in real-world conditions. It tracked a bright orange shuttlecock served by another machine while researchers monitored the speed, angles, and landing points of the shuttlecock.
The robot dog had to quickly cross the court to hit the shuttlecock hard enough to send it over the net and into the center.
After intensive training, the robot could track shuttlecocks and accurately return them at speeds of about 39 feet per second—nearly half the swing speed of an average amateur. ANYmal also adeptly adjusted its movement trajectory based on its distance from the shuttlecock.
Yuntao Ma was impressed by how coordinated the robot dog was in moving all 18 joints. This is particularly challenging since each joint learns independently, but executing a movement requires the synchronized effort of all of them.
The team also enthusiastically noted that ANYmal returned to the center of the court after each shot, just like human players do.
However, unfortunately, the robot dog did not account for the opponent’s movements. Researchers plan to teach ANYmal this in the next phase.
Ma believes this research will ultimately find applications beyond sports. It could assist, for example, in clearing debris during disaster recovery, he said. After all, ANYmal effectively combines visual perception of the situation with dynamic movements.
The results of the study were published in the journal Science Robotics.