How Tencent’s WeKick AI Dominated the Google Football Kaggle Competition

WeKick AI Wins the First Google Football Kaggle Competition

In the inaugural Google Football Kaggle contest, Tencent AI Lab’s WeKick AI achieved a total score of 1785.8, securing the champion title among more than 1,100 research teams worldwide.

The competition used the Google Research Football reinforcement‑learning environment, a 11‑vs‑11 setting where each player is controlled by an independent agent, demanding sophisticated team coordination and strategy.

Key Technical Innovations Behind the Victory

WeKick’s success stems from several core advances:

Migration of the full‑stack “Juewu” architecture to the football domain, incorporating customized framework improvements.

Integration of Generative Adversarial Imitation Learning (GAIL) with carefully designed reward shaping to emulate expert behavior.

Adoption of an asynchronous distributed reinforcement‑learning framework that enables flexible resource scaling during training.

Implementation of a multi‑style “League” training scheme, where a pool of specialized agents (each focusing on a distinct play style) are periodically pitted against a central model to avoid over‑fitting to a single style.

The “specialize‑then‑integrate” process first trains a competent base model (dribbling, passing, shooting), then creates multiple style‑specific agents, and finally consolidates them into a master model that can handle diverse opponent strategies.

Internal scoring shows the master model gains over 200 points beyond the base model and outperforms the strongest style‑specific agent by 80 points.

Broader Implications

The championship demonstrates that deep reinforcement learning methods developed for games like Go and MOBA can be transferred to more complex, multi‑agent environments such as football, marking a step toward general artificial intelligence.

Beyond gaming, the underlying research and algorithms are expected to benefit fields like agriculture, healthcare, and smart cities, unlocking substantial practical value.