TacticAI is a research system from Google DeepMind, developed with experts from Liverpool FC over a multi-year collaboration and published in Nature Communications on 19 March 2024. It focuses on corner kicks, one of football’s rare set-piece situations, and combines predictive and generative components to help coaches design and refine them.
The technical core is geometric deep learning with graph neural networks: each player is a node and the relations between them are edges, so the model reasons about the configuration of players rather than raw pixels. Because a football pitch is approximately symmetric, the team generated four reflected versions of each situation to squeeze more learning out of scarce data - only about ten corners are played per match. From a setup, TacticAI predicts the first player to receive the ball and the probability that the corner leads to a shot, and it can generate adjusted positions and movements that change the likely outcome.
The standout result was human evaluation: Liverpool experts preferred TacticAI’s suggestions 90 percent of the time over the setups actually seen in matches, and its top retrieved similar plays were judged relevant 63 percent of the time versus a 33 percent position-similarity baseline. It is one of the clearest demonstrations that learned tactical assistants can produce advice elite practitioners find genuinely useful, not just plausible.