Researchers at Switzerland's ETH Zurich have built a four-legged robot that can autonomously play badminton against human opponents using only onboard perception. The work was published in the journal Science Robotics, and footage released by the university shows the robot tracking and predicting the shuttlecock, positioning itself on four legs, and swinging a racket to return shots.
August 25, 2025 · ETH Zurich — Robotic Systems Lab
A Robot Dog Learned to Play Badminton — and Rallies With Humans
The quadruped ANYmal-D, fitted with a robotic arm, tracks, predicts, and returns shuttlecocks — coordinating 18 degrees of freedom under a single reinforcement-learning policy to sustain rallies of up to 10 shots.
18
degrees of freedom — legs + arm, one RL policy
10
consecutive shots in a rally vs. a human
12.06 m/s
peak racket swing speed
45°
fixed racket angle, optimal in simulation
FROM ONE PLATFORM TO A NEW SKILL
ANYmal began as an industrial inspection robot. This work extends the same legged platform into a dynamic, interactive sports task.
ORIGINAL ROLE
Industrial Inspection
Gas-leak detection and routine plant monitoring — slow, structured environments.
→
NEW DEMONSTRATION
Whole-Body Badminton
Perceive, predict, move, and swing — fast whole-body response under tight time limits.
THE PLAY LOOP — A SINGLE LEARNED POLICY
1 · Perceive
Stereo camera detects the shuttlecock; posture adjusts to keep it in view.
→
2 · Predict
Trajectory model forecasts where the shuttle will land.
→
3 · Move
Gait adapts to distance and time to reach the return position.
→
4 · Swing
Arm swings to return the shot — then the loop repeats.
WHAT IMPRESSED RESEARCHERS
18 DoF coordinated under one RL policy
All computation and perception run onboard
Tested in lab, a historic machine hall, and outdoors in wind
Active perception and gait adaptation emerged during training
REMAINING CHALLENGES
Some return failures remain — not perfectly human-level
Hardware limits: current draw, latency, system identification
A research prototype, not a commercial product
Generality to humanoids shown only in simulation so far
The takeaway: By fusing locomotion and manipulation into a single visuomotor policy — and closing the sim-to-real gap with a camera-noise model — the work points toward sports-training aids, dynamic-manipulation benchmarks, and human-robot interaction testbeds.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…