ETH Zürich trains robot to play badminton using whole-body reinforcement learning
lukas_m_ziegler · x · 2026-08-24
Researchers from ETH Zürich and Legged Robotics have trained a legged robot to play badminton, taking it beyond simple walking tasks. Instead of controlling limbs separately, the team used a unified system based on reinforcement learning to manage the robot’s entire body, coordinating movement and arm swings simultaneously.
Key technical highlights include:
- Prediction Model: The robot anticipates the shuttlecock's landing spot, similar to human players, to decide where to move next.
- Robust Vision: The robots were trained using real-world blurry or noisy visuals, ensuring they perform well even with imperfect sight outside the lab.
More from Embodied
- Robotics project "52 robots in 52 weeks" tests latest dog-like bot — _Stocko_ · 2026-08-24
- RTX 5090 Benchmark: Generating 12-Second 1080p Video with Minimax H3 Takes 35 Minutes — Grinderius · 2026-08-24
- Hands-on NVIDIA DGX Spark: Benchmarks and Tradeoffs for 24/7 Local Agents — JeremyNguyenPhD · 2026-08-24
- XPeng Robotics raises $900M, aims for 1,000 humanoids/month by 2026 — CyberRobooo · 2026-08-24
- Tsinghua's OVOW Reconstructs Interactive 4D Physical World from Single Video — jiqizhixin · 2026-08-24
- Opinion: US lags behind China in robotics, relying on parts and low volume — arian_ghashghai · 2026-08-24