Tsinghua Team Publishes Humanoid Robot Soccer Research in Science Robotics

In 2004, Zhao Mingguo led students to establish the Tsinghua Vulcan robot soccer team, which competed in RoboCup the following year. Early robots were only about 50 centimeters tall, walked slowly, and required spotters to prevent falls during matches. Over the next two decades, the Vulcan team competed almost every year, with robots growing to over a meter tall, learning to stand up autonomously, run quickly, and shoot powerfully.

The soccer field became a long-term testing ground for the team's vision, decision-making, motor control, and overall robot reliability. How many milliseconds slower the vision recognition was, how many centimeters off the foot placement was, and in which direction the robot tended to miss the ball were all directly revealed in real competition. The team then brought these issues back to the lab, modified algorithms, and continued validation in the next match—a cycle that persisted for over 20 years.

The Vulcan team's membership changed over generations. Cheng Hao, founder of Accelerated Evolution, was the third captain of the Vulcan team and founded 'Accelerated Evolution' in 2023, with several former team members joining. The paper's first author, Wang Yushi, comes from the new generation of the Vulcan team. From 2004 to the paper's publication in a top international journal in 2026, this is a 22-year technology relay.

The paper addresses a long-standing challenge: how can humanoid robots achieve see-and-react in real environments filled with noise, latency, and occlusion? Traditional approaches decompose the process into multiple modules, which can lead to error and latency accumulation, and policies trained in simulation may degrade in the real world due to lighting, noise, or occlusion.

Zhao Mingguo's team proposed a unified perception-motor reinforcement learning framework, training vision and motor control end-to-end under a single optimization objective. They first built a virtual perception system: real robots observed a soccer ball at various distances and angles, collecting about one hour of data, which was compared with motion capture ground truth to derive perception characteristics such as position noise, detection success rate, update frequency, and latency.

The vision system updates at approximately 25 Hz with an average latency of about 116 milliseconds. The policy reads the past 50 frames (about one second) of observations, compresses them into a 64-dimensional hidden state, and uses a decoder to aid visual denoising, allowing the robot to continue estimating the ball's position after brief occlusion. Training data includes about 76 seconds of human omnidirectional walking and 30 seconds of instep shooting motions, with adversarial motion prior (AMP) and mirror symmetry constraints. The final controller outputs joint position commands at 50 Hz.

Emergent behaviors appeared during training: when the ball approached the edge of the field of view, the robot actively turned its head and body; when the ball was not found near the boundary, it first oriented toward the center and then expanded its search; gait varied with distance, with a step cycle of about 0.7-0.8 seconds at long range and shortening to 0.3-0.4 seconds at close range; when the goal was behind, the policy generated a spinning hook shot, pivoting on the support foot.

Experimental data show that the raw visual ball position error one second before shooting was 0.344 meters, which dropped to 0.186 meters after estimation by the policy's internal state, a reduction of about 46%. In simulation, with a 0.3-second visual interruption, the ball contact success rate still exceeded 90% and the goal rate exceeded 50%. For a stationary ball, the learned policy typically took about 1.5 seconds from initiation to contact, while a traditional rule-based system took about 2-5 seconds, with a maximum time reduction of 64%.

In real-robot tests, the success rate in the front field was 80%-90%, and in the back field 60%-70%, with no falls during testing. For a rolling ball, the success rate still exceeded 50% when the ball speed was below 0.5 meters per second. In 2025, the Vulcan team won both the RoboCup Adult-Size Humanoid League and the World Humanoid Robot Games, scoring 76 goals and conceding only 11 across both competitions; in 2026, they defended their title in the RoboCup Large Size competition.

The learned policy can be deployed directly from simulation to the real robot, but the algorithm is only one part. Joint response consistency, control cycle stability, and timely output from sensors and computing units all affect performance, and any underlying error can be amplified during high-speed motion. The real-robot experiments in the paper used the Accelerated Evolution humanoid platform without special modifications, with ball information from a head-mounted camera and perception and control running entirely on onboard computing.

The relationship between Accelerated Evolution and the Vulcan team extends from talent to technology platform. In July of this year, Accelerated Evolution released its new flagship platform, Booster T2, with the professional version equipped with an NVIDIA Thor chip, delivering 2070 TFLOPS of edge computing power, claimed to be the most powerful computing platform in the bipedal humanoid field. The accompanying Booster Studio integrates simulation training, algorithm development, and real-robot deployment into a single development environment.

On August 19, Booster T2 debuted at the World Robot Conference. Three days later, 80 Booster T2 units performed centerless autonomous coordination at the opening ceremony of the Robot Games, writing 'BEIJING', the emblem, and event icons without remote control or commands. 80 independently deciding 'brains' achieved coordination simultaneously, with a total computing power of about 170,000 TFLOPS. From early robots that were 50 centimeters tall and needed human support, to Booster T2 autonomously playing soccer on a thousand-square-meter field and writing in formation at the opening ceremony, humanoid robots are moving from 'can move' to 'can be used'.