A team of engineers from the Robotic Systems Lab in ETH Zurich has introduced a four-legged robot capable of playing badminton with humans. Named ANYmal-D, the robot was trained using reinforcement learning to anticipate and respond to the shuttlecock during a rally.

The project, recently published in Science Robotics, explores how legged robots can operate in dynamic, fast-paced scenarios, an area not often tested with traditional machine control systems. Rather than focusing on humanoid mimicry, the researchers opted for a quadrupedal design to prioritize stability and responsiveness.

How It Works

To operate effectively on a badminton court, the robot was equipped with:

1. A stereo vision system for depth perception and shuttlecock tracking
2. A robotic arm to return shots
3. A reinforcement learning controller that helped it position itself on the court, anticipate incoming shots, and swing in response

Additionally, the researchers introduced a "perception noise model", which allowed the robot to handle real-world visual inconsistencies by comparing live sensor data with data from its training environment. This helped the robot maintain performance when environmental conditions shifted slightly such as lighting changes or motion blur from fast shuttlecock movements.

Key Observations

Testing showed that ANYmal-D could sustain rallies with human opponents for up to 10 consecutive shots. While it does not play at a competitive level, the robot successfully demonstrates that legged machines can participate in interactive and real-time tasks that involve quick positioning and coordination.

According to the research team, the quadruped's design gave it greater flexibility compared to bipedal systems. Its ability to tilt its base for visual tracking and rapidly shift positions gave it enough agility to keep pace with a human opponent across a small court.

Implications

This experiment doesn't suggest robots are ready to compete in sports, but it offers valuable insights for robotics applications that require real-time motion planning and coordination such as search-and-rescue, warehouse automation, or rehabilitation support.

The study underlines an important direction for robot learning systems: rather than perfecting individual motions in isolation, machines can be trained to adapt dynamically to moving objects and unpredictable scenarios without relying on hard-coded rules.