Monte Carlo Pass Search: Using Trajectory Generation for 3D Counterfactual Pass Evaluation in Football
Researchers introduce Monte Carlo Pass Search (MCPS), a novel AI system that evaluates football passes by simulating counterfactual scenarios using trajectory generation and value prediction models. The work combines existing machine learning techniques with a new public Bundesliga dataset featuring 3D ball tracking, enabling distribution-aware analysis of pass execution quality and decision-making.
This research represents a meaningful advancement in sports analytics by applying sophisticated machine learning techniques to football pass evaluation. Rather than relying on traditional metrics, the team developed a system that simulates alternative pass outcomes through trajectory generation, treating pass evaluation as a Monte Carlo sampling problem. The approach integrates three key components: a value model measuring possession quality, a world model predicting multi-agent behavior with ball interactions, and a policy for generating pass variations with realistic execution noise.
The work builds on growing interest in applying advanced AI methods to sports, particularly in understanding player decision-making and match dynamics. By releasing a public high-fidelity tracking dataset from the Bundesliga with 3D ball trajectories, the researchers contribute valuable infrastructure for the sports analytics community. The adaptation of SMART, a trajectory generation model from autonomous driving, demonstrates cross-domain AI transfer and highlights how techniques from one field can solve problems in another.
The distribution-aware attribution approach offers practical value for teams and analysts seeking to understand not just whether passes succeeded, but how execution uncertainty affected outcomes. Mean-based and percentile-based scoring metrics provide complementary perspectives on pass quality. The release of model checkpoints and code enables broader adoption and validation.
Looking forward, similar counterfactual evaluation methods could extend to other tactical decisions in football and potentially other sports. The success of cross-domain model adaptation suggests opportunities for applying autonomous driving and robotics research to sports analytics more broadly.
- →MCPS uses Monte Carlo sampling to simulate pass variants with realistic execution noise, enabling distribution-aware evaluation rather than binary success/failure assessment
- →The system combines trajectory generation from autonomous driving with sports-specific value models, demonstrating effective cross-domain machine learning transfer
- →A new public Bundesliga dataset with 3D ball tracking provides essential infrastructure for advancing football analytics research
- →Distribution-based scoring metrics offer complementary mean and percentile perspectives on pass quality, providing richer insights than traditional statistics
- →Open-sourced code and model checkpoints enable broader adoption and validation within the sports analytics community