Article

Football AI searches for better corners before the kick

For football fans and analysts, a simulation model compares crosses and starting positions—and shows why estimated gains are not extra goals.

Editorial illustration for Football AI searches for better corners before the kick: a geometric block represents a model release. Not documentary evidence.

A football AI study released in Hudl’s November 19 research collection suggests that changing the cross and starting positions offers more simulated attacking value than designing extra runs. For fans and analysts, the promise is a way to compare corners that were never taken—not evidence that a club will score more goals.

The revealing detail is how little ground the attackers covered while the ball was airborne. In Andrew Kang’s study, they moved a median 2.2 metres between the kick and the ball’s descent. That brief window helps explain why the model found more room for improvement in where players started than in additional choreography.

Kang presented the work at Hudl Performance Insights on November 13; the paper collection and its arXiv submission followed on November 19. The experiments cover 249 lofted deliveries—170 corners and 79 free kicks—from 30 matches among six Serie A clubs in the 2025/26 season.

What can AI compare that a replay cannot?

A clearance answers what happened to one ball. It does not settle whether a different delivery, with the same players available, offered a better chance. The analytical distinction is between judging an outcome and judging the options that preceded it.

Kang’s method, Monte Carlo Pass Search, generates alternative deliveries and starting positions. It repeatedly simulates each option, varying the execution of the kick and the players’ responses. Defenders move, attackers pursue the ball and the goalkeeper decides whether to come for it. A candidate is judged across many possible short sequences, not one perfect execution.

The simulation stops at the first touch. Another model estimates the chance of an attacking shot within the next three seconds and the quality of that shot. That quality is expressed as expected goals, or xG: an estimate of how likely a chance is to become a goal. The combined score measures a brief attacking opportunity, not the outcome of a whole possession.

This extends Kang and Priya Narasimhan’s June research on alternative open-play passes. It also enters an established field: Google DeepMind’s TacticAI work with Liverpool FC, announced in March 2024, already explored different corner setups. DeepMind’s reported 90% figure measured experts’ preference for suggestions, not a scoring rate. It does not verify the new study’s results.

Why did starting positions offer more than extra runs?

In a separate experiment that also modeled defensive setups, searching the attackers’ starting positions added an average 0.0033 xG per delivery beyond the best searched ball. Once both ball and setup had been chosen, designing runs added another 0.0007 xG per delivery. The extra run value was smaller for corners than for free kicks.

The 2.2-metre movement finding makes that result easier to picture: there is limited time to rearrange an attack after the kick. Starting nearer a useful area can matter before the first purposeful stride. That is an interpretation of this experiment, not a general argument against clever movement.

Short routines are excluded, and the movement window does not measure everything players do before the kick. It cannot establish that decoys or rehearsed runs are unimportant. Nor does “additional value after other choices were optimized” tell us the total value of running.

Readers interested in how machines explore coordinated choices can also see how IBMARL helps AI teammates learn beyond imitation. That is a different simulated task, not evidence that these football suggestions will transfer to a pitch.

What does the reported 28% gain mean?

On average, the paper’s delivery-search comparison assigns the observed deliveries 0.0307 xG each and the best searched alternatives 0.0085 more. A calculation from those reported inputs gives 0.0392 xG per delivery, with the increase equal to 0.0085 ÷ 0.0307, or about 28%. This assumes the paper’s scoring model is an adequate yardstick; it is not 28% more actual goals.

Restricting the search to deliveries resembling real ones leaves a reported gain of 0.0079 xG. That realism check reduces the chance of recommending an implausible ball, but it does not prove a particular taker could execute it. These delivery-only results and the setup experiment use different evaluation conditions; their gains should not be added together.

Kang also re-evaluates selected options with fresh simulations and separately trained scoring models, and keeps evaluated matches out of the models’ training data. Those checks make it harder for the search to win merely by finding options its own scorer overvalues. The “independent judges” are models, however—not independent researchers or a coaching trial.

There is a related evaluation question in JioHotstar’s use of sports experts to check AI search: who checks the checker? Testing a cricket answer is different from testing a corner routine, but agreement between machines is not the same as evidence from the task itself.

Which suggestions deserve a closer look?

The most useful output may be an area to investigate, not a dot to hit. Separately trained pipelines disagreed on the preferred delivery target by a median 2.7 metres. That supports discussing delivery zones with an analyst rather than issuing precise instructions from one model’s favourite point.

Starting-position advice carries another cost the model does not count. Moving an extra attacker forward can improve the immediate attacking score while leaving fewer players to stop a counterattack. The paper explicitly omits that later defensive risk, so it does not settle whether the move helps the team overall.

It also omits fouls and individual heading, jumping and goalkeeping ability. Marking and goalkeeper responses are imperfect, and the six-club sample does not establish the same result across other leagues or types of football. Those limits affect which alternative is preferable, not just how confidently to describe it.

This remains research, not an announced coaching product. The public repository for the earlier MCPS project requires prepared data and excludes large model and dataset files from Git. It does not establish access to the complete November system or a verified operating price.

For an analyst, the practical next step is to examine candidate delivery zones and starting setups alongside execution ability, marking and the players left back. Any proposed change then needs evaluation beyond this simulator. For a fan, the reward is a better question after a clearance: what other choice was available before the ball left the taker’s foot?