Robot Smashes It: ANYmal Learns to Play Badminton

At ETH Zurich, researchers have trained the legged robot ANYmal to track, move, and return shuttlecocks in real time. Using a smart control system that syncs vision, locomotion, and striking, ANYmal predicts trajectories and positions itself for a perfect shot—showcasing the future of agile, perceptive robots in dynamic environments.

Transcript

00:00:03 ANYmal is a quadruped robot. So it has four legs. We have eyes, and the robot mimics that. The robot has two cameras, and when a person hits the shuttle, it looks for the color of the shuttle

00:00:17 and tries to find where the color is. And so the robot is just, in a mathematical way, calculating where it is, and where it's going to land. And then we have a controller— something that moves the motors of the robot, that actually moves

00:00:32 and executes the shot itself. When you're running for a shuttle, it's easy to lose track of what your limbs are doing. And all of this together is very challenging. The control of the robot is based on reinforcement learning.

00:00:44 So reinforcement learning means taking a robot and throwing it into a simulator. And in the simulator, it can do thousands and thousands of iterations of trial and error, slowly learning what is the best strategy. Sports are very challenging,

00:01:00 because they require a lot of fast moving dynamics. It pushes the limits of the system, and by pushing the limits, we make many iterative improvements on how we control the robot, and on the system itself. These improvements then feed into other projects.

00:01:16 For now, it's still friendly matches. I would rate it like a seven-year-old kid playing with their parent. So if someone tried to actually beat it, the robot would have no chance. I think that's something we’re working towards—

00:01:30 so you can actually play against it seriously.