Tsinghua Team Publishes in Science Robotics: Humanoid Robots Master Reactive Soccer Skills

机器之心 · wechat · 2026-08-24

Tsinghua University, in collaboration with ByteDance, published a paper in Science Robotics presenting a vision-driven framework for learning reactive soccer skills in humanoid robots, addressing challenges in real-time perception and control within noisy, dynamic environments.

Methodology: The team uses end-to-end reinforcement learning to unify visual perception and motion control. An encoder-decoder architecture handles visual latency and occlusion, while Adversarial Motion Priors (AMP) are introduced to enhance natural movement.
Performance: In simulation, the robot maintains a >90% ball-touch success rate even with a 0.3s visual interruption. Real-world tests show a time-to-touch of just 1.5 seconds—64% faster than traditional rule-based systems—with a success rate of 80-90% in the front field.
Hardware: Experiments were conducted on humanoid robots from Accelerate Evolution (加速进化), utilizing onboard vision and compute, demonstrating the capability of domestic platforms to support state-of-the-art embodied intelligence research. This technology helped the Tsinghua Husky team win the RoboCup championship.

Original post →

Tsinghua Team Publishes in Science Robotics: Humanoid Robots Master Reactive Soccer Skills

More from Embodied