Tsinghua and Booster Robotics Solve 'Blurred Vision' Problem: Zero-Shot Sim-to-Real Soccer Skills for Humanoid Robots on Science Robotics Cover

Forum topic · QianXun · 2026-08-24

Summary

On August 19, Science Robotics featured on its Humanoid Robots special-issue cover a Tsinghua University study, 'Learning Vision-Driven Reactive Soccer Skills for Humanoid Robots,' developed with Booster Robotics and tested on the Booster T1 humanoid. For two decades, humanoid soccer has been limited by the classic perception-planning-control pipeline: modules operate in isolation, serial communication causes decision lag, and simulation-to-reality gaps cripple real-world performance, especially when onboard cameras blur during running or lose sight of the ball. The paper addresses all three problems with an end-to-end framework that jointly optimizes perception and control under a single reinforcement-learning objective; it extends adversarial motion priors (AMP) from simulation to real dynamic vision for zero-shot sim-to-real deployment; and it uses a virtual perception system with an encoder-decoder architecture that compresses 50 frames (1 second) of history into a 64-dimensional latent state to reconstruct ball position, inferring trajectories even during brief occlusion. Trained entirely in simulation, the policy reduced ball-position estimation error by 46%, cut kicking time by 64%, and achieved roughly 90% success on forward-field kicks with no real-world fine-tuning.

Background

On August 19, the *Science Robotics* "Humanoid Robots" special issue featured on its cover a paper by Professor Zhao Mingguo's team from Tsinghua University's Department of Automation: *Learning Vision-Driven Reactive Soccer Skills for Humanoid Robots*. The robot platform is the Booster T1, developed by Booster Robotics.

Why This Problem Was Stuck for 20 Years

The hardest part of humanoid soccer isn't running, balance, or shot power—it's that the robot "can't see clearly" while moving: the camera shakes violently during running, sunlight makes the ball blend into the field, and a single collision can remove the target from view. For humans this means blurry vision; for a robot, it means the decision system completely fails.

For twenty years, the traditional architecture hasn't escaped the same pipeline: a vision module "sees" → a planning module "thinks" → a control module "acts." This works in the lab but hits three fatal walls in the real world:

Module separation: the robot moves even when the perceived position is wrong
Serial communication: decisions are always lagging behind reality
Sim-to-real gap: policies trained in perfect simulation fail on real hardware

Three Breakthroughs

1. End-to-end perception-action coupling. Perception and control are jointly optimized under a single reinforcement-learning objective—the robot "prepares to act" at the same moment it "sees," eliminating serial latency.

2. Adversarial motion priors (AMP) extended from simulation to real dynamic visual scenes for the first time. This enables zero-shot sim-to-real deployment with no real-world fine-tuning.

3. Virtual perception system with an encoder-decoder architecture. The encoder takes the past 50 frames (1 second) of observations, compresses them into a 64-dimensional latent state, and the decoder reconstructs the ball's position—inferring its trajectory even when the ball is briefly occluded.

Real-World Performance

All policies were trained only in simulation and deployed on real hardware with zero adaptation:

Ball position estimation error: reduced by 46%
Kicking time: reduced by 64%
Forward-field kick success rate: ~90%

Why Booster Robotics

Booster's platform has become the de facto standard for robot soccer research teams worldwide:

RoboCup 2026 (July) humanoid league: 38 international teams (about 68%) chose Booster, which swept all gold medals in the small, medium, and large size classes
World Humanoid Robot Games (August) soccer events: 56 teams (92%) chose Booster
First author Wang Yushi is captain of Tsinghua's Vulcan (Huo Shen) team, which won the RoboCup humanoid league in both 2025 and 2026

The newly released Booster T2 uses the NVIDIA Thor chip with 2070 TFLOPS of onboard compute (currently the highest among bipedal humanoids), stands 1.4 m tall with 31 degrees of freedom and a 10 kg dual-arm payload. Booster Studio, billed as the first IDE for embodied intelligence, unifies simulation, policy training, and real-robot deployment—exactly the engineering foundation behind the paper's zero-shot sim-to-real pipeline.

Takeaway

The significance of this paper goes beyond "pretty soccer." It pushes embodied intelligence from the "lab pipeline" toward "unified perception-motor control in the real world": the first real-robot AMP deployment, zero-shot sim-to-real, and trajectory inference under occlusion together point to one conclusion—humanoid robots are a step closer to acting reliably in messy real environments. And with 68% platform share at RoboCup and 92% at the World Humanoid Robot Games, Booster is turning a "research standard" into an "industry fact."

References

1. *Science Robotics* cover paper, 8/19/2026, DOI 10.1126/scirobotics.aed1152 2. China Daily report, 8/21/2026 3. Booster Robotics official press release, 8/20/2026 4. Tsinghua Department of Automation Weibo, 8/24/2026 5. Olympia Journal / Newsfile, 8/20/2026