Sports

From PRIMO.ai
Jump to navigation Jump to search

Youtube search... ...Google search

Artificial intelligence (AI) is changing how athletes train, coaches develop strategies, officials make decisions, teams evaluate talent, and fans experience sports. By combining machine learning, computer vision, statistical modeling, wearable sensors, and large-scale computing, sports organizations can extract useful information from movements and events that were previously difficult or impossible to measure.

Traditional sports analytics relied primarily on box scores, manually recorded events, historical averages, and human observation. Modern AI systems can analyze detailed player and ball trajectories, reconstruct body movements in three dimensions, recognize tactical formations, estimate the probability of future events, and suggest adjustments to training or game strategy.

The most important change is the movement from analyzing what happened to understanding how events unfolded and exploring what might happen under different conditions.

Applications now span professional and amateur sports, including:

  • Player performance analysis: Measuring speed, acceleration, positioning, technique, decision-making, and movement efficiency.
  • Game strategy: Evaluating formations, matchups, substitutions, passing options, and alternative tactical decisions.
  • Biomechanics: Reconstructing skeletal movement to study athletic technique and mechanical loading.
  • Injury risk assessment: Combining workload, movement, physiological, and medical information to identify possible risk factors.
  • Sports prediction: Estimating match outcomes, player performance, and the probability of specific events.
  • Automated officiating: Supporting referees and judges with optical tracking, ball sensors, and computer vision.
  • Scouting and recruitment: Comparing athletes across leagues, positions, playing styles, and competitive environments.
  • Equipment engineering: Using simulation and optimization to investigate golf clubs, bats, footwear, protective equipment, and other sporting products.
  • Fan engagement: Generating highlights, advanced statistics, interactive explanations, and personalized sports content.

Important organizations advancing the field include Google DeepMind, FIFA, Major League Baseball (MLB), the National Football League (NFL), the National Basketball Association (NBA), Amazon Web Services (AWS), Sony Hawk-Eye Innovations, Fujitsu, and academic research groups associated with the SoccerNet project.

AI is most effective when it supplements human expertise. Statistical prediction is not the same as causal explanation, and a computer-generated recommendation does not automatically establish that an intervention will improve performance.


Artificial intelligence is transforming sports from the retrospective analysis of statistics into the continuous measurement, modeling, prediction, and optimization of athletic performance. Its greatest value comes from combining computational insight with the judgment of athletes, coaches, medical professionals, and officials.



Mechanisms

Sports AI systems convert observations into measurements, measurements into mathematical representations, and mathematical representations into predictions or recommendations.

Most advanced implementations contain several connected components.

Data Collection and Sensor Systems

The quality of an AI model depends heavily on the underlying measurements.

Optical tracking

Stadium cameras capture images of athletes, equipment, and playing surfaces. Computer vision algorithms detect objects and estimate their locations across successive video frames.

Multi-camera installations can reconstruct three-dimensional positions through geometric calibration and triangulation.

Typical measurements include:

  • Player and ball coordinates.
  • Movement speed and acceleration.
  • Direction changes and body orientation.
  • Distances between opponents and teammates.
  • Body-joint positions and limb angles.
  • Passing, shooting, and collision events.

Broadcast video can also be analyzed, although moving cameras, changing zoom levels, occlusion, and players outside the camera's field of view create additional challenges.

Wearable sensors

Wearable devices can include:

  • Global navigation satellite system (GNSS) receivers.
  • Accelerometers.
  • Gyroscopes.
  • Inertial measurement units (IMUs).
  • Heart-rate monitors.
  • Positioning tags using radio-frequency or ultra-wideband technology.

These systems provide different kinds of information.

A satellite receiver estimates outdoor position. An accelerometer measures specific force. A gyroscope measures angular velocity. An IMU combines multiple motion measurements and may use sensor-fusion algorithms to estimate orientation.

Heart-rate measurements provide physiological information, although heart rate alone does not determine fatigue, recovery, or injury risk.

Instrumented sports equipment

Sensors embedded in balls or other equipment measure movement characteristics that may be difficult to determine from ordinary video.

FIFA's connected-ball technology, for example, uses an IMU operating at approximately 500 measurements per second to help identify the moment of ball contact.

The NFL uses radio-frequency tracking technology to measure player movement and ball-related information.

Medical and training records

Sports scientists may also use:

  • Previous injury records.
  • Training-session attendance.
  • Exercise volume and intensity.
  • Strength assessments.
  • Recovery measurements.
  • Sleep and wellness reports.
  • Match participation and playing time.

Because these datasets contain sensitive information, collection and analysis require appropriate consent, security, access controls, and governance.

Computer Vision and Motion Reconstruction

A typical vision-based sports analytics pipeline follows several stages.

  1. Video acquisition: Cameras record the event, preferably with accurate timestamps and known calibration parameters.
  2. Object detection: Neural networks locate players, officials, balls, and relevant equipment.
  3. Tracking: Algorithms associate detections across frames so individual identities remain consistent.
  4. Camera calibration: Image coordinates are transformed into field or court coordinates.
  5. Pose estimation: Models locate body landmarks such as shoulders, elbows, hips, knees, and ankles.
  6. Event recognition: Temporal models identify actions such as passes, shots, sprints, tackles, and jumps.
  7. Analytical modeling: Tracking data is converted into tactical, biomechanical, or statistical features.
  8. Decision support: Results are presented through dashboards, annotated video, alerts, or recommendations.

Object detection and tracking are related but different problems.

Detection answers, "Where is the player in this frame?"

Tracking answers, "Is this the same player observed in earlier frames?"

Tracking systems must handle players crossing paths, temporary disappearance, jersey similarities, substitutions, and rapid changes in direction.

Modern tracking approaches combine visual appearance, predicted motion, and contextual information. Examples include Kalman filtering, learned motion models, and neural networks that associate detections across time.

In football research, the SoccerNet Game State Reconstruction task evaluates whether a system can recover player identity, team affiliation, playing role, and pitch coordinates from broadcast footage.

SoccerNet Game State Reconstruction | SoccerNet Research benchmark and evaluation framework.

Machine Learning and Predictive Models

Once movements and events have been measured, AI systems can learn statistical relationships.

Supervised learning

A model learns from examples containing input data and known outcomes.

Applications include:

  • Predicting whether a pass will be completed.
  • Estimating shot or scoring probability.
  • Classifying defensive coverage.
  • Identifying sporting actions.
  • Estimating future player performance.

Algorithms range from logistic regression and gradient-boosted trees to neural networks and transformers.

Unsupervised learning

Clustering and dimensionality-reduction methods identify patterns without requiring predetermined outcome labels.

Examples include finding similar player movement profiles, identifying recurring tactical formations, or discovering distinct styles of play.

Graph neural networks

Team sports naturally contain interacting entities.

A graph can represent each player as a node and relationships between players as edges. Node features may include location, velocity, team identity, and orientation. Edges can represent proximity, passing opportunities, defensive assignments, or other relationships.

Graph neural networks update player representations by exchanging information across these connections.

This approach is especially useful when the behavior of one athlete depends on the positioning and intentions of several others.

Transformers and temporal models

Transformers use attention mechanisms to identify relationships among players and events across time.

They can process movement sequences and capture relationships that change during a play.

Other temporal approaches include recurrent neural networks, temporal convolutional networks, and state-space models.

Generative models

Generative AI can propose alternative tactical arrangements, reconstruct missing tracking information, create hypothetical movement sequences, or produce natural-language explanations.

These outputs must be evaluated against physical constraints, sporting rules, and real match data.

A plausible simulated play is not proof that the same outcome would occur in competition.

From Prediction to Intervention

A predictive system estimates an outcome conditional on available information.

For example, a model might estimate that a pass has a 70% probability of completion given the observed player positions.

A recommendation system addresses a different question: Would moving a player, changing a training plan, or choosing another tactical option improve the outcome?

This is more difficult because sports decisions are influenced by factors that may not be measured.

The distinction is important:

  • Prediction: What is likely to happen?
  • Explanation: Which observed variables are associated with the outcome?
  • Causal inference: What would change because of a specific intervention?
  • Optimization: Which permitted decision performs best under stated objectives and constraints?

A model can predict outcomes accurately while providing unreliable recommendations about changing those outcomes.

Controlled experiments, prospective evaluations, and carefully designed causal studies are needed to test whether AI-assisted interventions work.

Player Performance Analysis

Player performance analysis is one of the most established uses of AI in sports.

Traditional statistics summarize completed actions. Tracking systems add information about the movement and positioning that preceded those actions.

Movement and Workload

AI systems can evaluate:

  • Total distance traveled.
  • Distance covered at different speed thresholds.
  • Peak speed.
  • Acceleration and deceleration events.
  • Movement asymmetry.
  • Sprint frequency.
  • Positional changes.
  • Recovery between high-intensity efforts.
  • Repeated movement patterns.

These measurements can help coaches compare training demands with competitive demands.

However, external workload does not necessarily describe an athlete's internal physiological response.

Two athletes completing identical running sessions may experience different levels of fatigue because of differences in fitness, recovery, injury history, and other conditions.

Technique and Skill Development

Computer vision can identify movement landmarks and compare technique across repeated performances.

Examples include:

  • Baseball pitching mechanics.
  • Golf swing geometry.
  • Tennis serving motion.
  • Sprinting posture.
  • Basketball shooting technique.
  • Gymnastics landing positions.
  • Weightlifting movement patterns.

AI can provide feedback about observable differences in technique, but identifying an unusual movement is not equivalent to identifying a defect.

Elite athletes often develop individual movement strategies that differ from average biomechanical patterns.

Coaches should interpret model outputs alongside performance outcomes, athlete characteristics, and medical assessments.

Scouting and Talent Identification

Machine learning can help identify athletes with particular performance profiles.

Scouting models may combine:

  • Competition statistics.
  • Age and development history.
  • Physical characteristics.
  • Position-specific measurements.
  • Quality of opposition.
  • Tactical role.
  • Tracking-derived movement features.

A model should account for differences between leagues and playing environments.

A player's statistical performance may change substantially when moving to a different team, league, position, or tactical system.

AI can narrow a scouting search, but it cannot reliably replace contextual assessment of adaptability, health, motivation, and other difficult-to-measure factors.

Dutch sports analytics company SciSports uses emerging technology to innovate on the pitch | SAS An established example of data-driven player analysis and football scouting.

Game Strategy and Decision-Making

AI can help coaches analyze how individual decisions interact with team behavior.

Applications include:

  • Lineup and substitution analysis.
  • Opponent tendencies.
  • Formation recognition.
  • Defensive coverage classification.
  • Passing lane evaluation.
  • Set-piece optimization.
  • Expected possession value.
  • Late-game decision support.
  • Scenario simulation.

Real-time applications must account for processing latency, changing game states, communication constraints, and applicable competition rules.

Expected Value and Decision Quality

Sports strategy often involves comparing actions with uncertain outcomes.

For example, a football model may estimate the probability that a pass reaches a teammate and the expected value of the possession afterward.

A basketball model may compare an open three-point attempt with a contested shot near the basket.

An expected-value framework can be written as:

Expected value = sum of (outcome probability × outcome value)

The result depends on the model's assumptions and how outcome value is defined.

Maximizing immediate scoring probability is not always the same as maximizing the probability of winning the match.

Game state, time remaining, athlete fatigue, and opponent behavior can change the preferred decision.

TacticAI: AI-Assisted Football Tactics

Google DeepMind and Liverpool FC developed TacticAI, a research system for analyzing and proposing football corner-kick strategies.

The work was published in Nature Communications in March 2024.

The system represents players and their relationships using geometric deep learning and graph neural networks.

Its predictive component estimates outcomes such as likely ball receivers and whether a shot may result.

Its generative component proposes changes to player positioning.

The researchers evaluated the system using historical corner-kick data and assessments by Liverpool FC experts.

In a blind expert evaluation, AI-generated suggestions were preferred over the comparison tactics 90% of the time.

This result demonstrates that the model could generate recommendations judged useful by specialists in the study.

It does not establish that using those recommendations produces a 90% improvement in goal scoring or match results.

The research is important because it moves beyond descriptive analytics toward the generation and assessment of alternative tactical arrangements.

TacticAI: An AI Assistant for Football Tactics | Wang, Veličković et al., Nature Communications, 2024 Peer-reviewed description of the model architecture, predictions, and expert evaluation.

TacticAI Research Overview | Google DeepMind, March 2024 Accessible explanation of the collaboration and practical applications.

Injury Risk, Recovery, and Athlete Health

AI-based injury prediction is a promising but methodologically challenging field.

Researchers combine training exposure, previous injuries, physiological data, biomechanical measurements, and other variables to estimate injury probabilities.

Potential applications include:

  • Identifying unusually high workloads.
  • Detecting changes in movement patterns.
  • Supporting return-to-play discussions.
  • Monitoring rehabilitation progress.
  • Studying recurring injury risk factors.
  • Comparing recovery patterns across training periods.

A useful system must do more than identify statistical associations.

It should be tested prospectively and demonstrate that its predictions support better decisions.

Limitations of Injury Prediction

Injuries are comparatively infrequent events, and many contributing factors are difficult to measure.

A model may achieve high overall accuracy by predicting that most athletes will not be injured.

For example, if injuries occur in only 5% of observations, a model that always predicts no injury would obtain 95% accuracy while failing to identify every injury event.

Evaluation should therefore consider:

  • Sensitivity and specificity.
  • Precision and recall.
  • False-positive and false-negative rates.
  • Calibration of predicted probabilities.
  • Class imbalance.
  • External validation.
  • Performance across different populations.
  • Prospective clinical usefulness.

A 2026 systematic review of 18 empirical sports injury prediction studies found that only one study had a low overall risk of bias under the review's assessment.

The authors highlighted limited external validation, inconsistent methodology, and underrepresentation of female athletes.

A separate 2026 systematic review focused on football injury prediction also identified substantial methodological limitations and warned against interpreting reported accuracy as sufficient evidence for routine clinical implementation.

The practical conclusion is that AI-generated injury scores should remain supporting information rather than independent medical judgments.

From Data to Prevention: A Systematic Review of Artificial Intelligence Applications in Sports Injury Prediction | 2026 Review of model development, validation quality, and research limitations.

Machine Learning Models for Injury Risk Prediction in Football Players | 2026 Systematic review of predictors, reported performance, and practical limitations.

Football and Soccer Analytics

Football is particularly suitable for spatial AI research because team behavior depends on continuously changing relationships among players and the ball.

Tactical Tracking

Tracking models can help researchers examine:

  • Team compactness.
  • Defensive line height.
  • Passing networks.
  • Pressing intensity.
  • Off-ball movement.
  • Available passing lanes.
  • Spatial control.
  • Set-piece organization.
  • Transition patterns.

Spatial control models estimate which team or player can reach a region of the pitch under assumed movement capabilities.

Different models may produce different results depending on acceleration assumptions, reaction time, player orientation, and the treatment of obstacles.

Tracking data can support tactical interpretation, but it does not directly reveal player intentions or coaching instructions.

SoccerNet Research

SoccerNet is an academic research initiative providing datasets, benchmarks, and evaluation tasks for football video understanding.

Its work includes:

  • Action spotting.
  • Ball action recognition.
  • Camera calibration.
  • Player re-identification.
  • Multi-object tracking.
  • Jersey number recognition.
  • Game state reconstruction.
  • Multi-view foul recognition.
  • Monocular depth estimation.

The 2025 SoccerNet challenges emphasized four areas: team ball action spotting, monocular depth estimation, multi-view foul recognition, and game state reconstruction.

These challenges are useful because they establish standardized datasets and evaluation protocols.

They also reveal where current systems still struggle, including the reliable reconstruction of events from incomplete broadcast views.

SoccerNet 2025 Challenges Results | Giancola, Cioppa et al., 2025 Research report describing the datasets, tasks, competing methods, and benchmark results.

SoccerNet Development Resources | SoccerNet Research code, datasets, and task documentation.

AI-Supported Football Officiating

FIFA introduced semi-automated offside technology at the 2022 FIFA World Cup.

The technology combines calibrated stadium cameras, player tracking, and ball-contact information to support offside assessments.

FIFA's advanced system uses 16 dedicated cameras, typically capturing 29 body landmarks per player at 50 measurements per second.

Where a connected ball is available, its IMU provides additional information about the timing of ball contact.

The system produces proposed positional offside decisions and supports review by match officials.

It does not independently resolve every judgment involved in offside, particularly questions about interference with an opponent.

An automated measurement of player position should not be confused with complete automation of the Laws of the Game.

FIFA to Install AI to Help Make Accurate Offside Decisions | Juana Summers, NPR, July 2022 Historical coverage of the technology's introduction.

Semi-Automated Offside Technology | FIFA Technical explanation, testing, and the limits of automated decision support.

American Football Analytics

The NFL's Next Gen Stats system combines sensor-based player tracking with machine learning.

According to NFL Football Operations, the system uses stadium radio receivers and tracking tags installed in player equipment and other game objects.

Player positions are sampled approximately 10 times per second.

These measurements support features such as:

  • Player speed and acceleration.
  • Distance traveled.
  • Route detection.
  • Defensive coverage analysis.
  • Completion probability.
  • Expected rushing yards.
  • Matchup evaluation.
  • Movement-based performance metrics.

Advanced Defensive Analysis

In 2025, the NFL introduced updated models for interpreting defensive coverage responsibilities.

The models analyze spatial and temporal relationships among players to estimate coverage assignments, matchups, and defensive disguises.

The league described using transformer-based models to process tracking information across successive frames.

This allows analysts to distinguish a defender's location from the role that defender may be performing.

A defender positioned near a receiver is not necessarily responsible for covering that receiver.

The NFL also updated its Completion Probability model in 2025, incorporating additional features to represent contested catches and the geometry of the passing path.

These developments extend the original Next Gen Stats approach from tracking observable movement toward inferring tactical roles.

NFL Next Gen Stats | NFL Football Operations Official description of the tracking system and applications.

Next Gen Stats: Advanced Metrics for the 2025 Season | NFL Explanation of coverage models, defensive disguises, and updated completion probability.

The NFL and Amazon Are Using AI to Invent New Football Stats | Malak Saleh, Engadget Background on the partnership and the development of movement-based statistics.

Basketball Analytics

Basketball analytics increasingly relies on continuous player tracking rather than exclusively on traditional play-by-play statistics.

The NBA began deploying Sony Hawk-Eye's three-dimensional optical tracking system during the 2023-24 season.

The resulting positional and skeletal data creates opportunities to examine spacing, defense, shot difficulty, and ball-handling.

Three-Dimensional Tracking

The NBA's system captures body-pose information and the position of the ball.

According to the league's October 2026 technical descriptions, the system measures 29 anatomical points per player at 60 measurements per second.

This enables models to observe how a player's body orientation and movement relate to teammates, defenders, and the ball.

New NBA Metrics

In October 2026, the NBA introduced technical descriptions of new machine-learning-based basketball statistics.

Handle Score

Handle Score attempts to quantify the effectiveness and complexity of ball-handling sequences.

The NBA describes a model that combines full movement sequences with characteristics such as dribble moves and defender separation.

The model was developed using more than 10,000 expert pairwise comparisons across more than 1,000 plays.

The training approach is important because it uses human preference judgments rather than relying entirely on manually specified scoring rules.

However, a model trained to reproduce expert preferences does not establish a universal or entirely objective definition of basketball skill.

Zone defense detection

The NBA also described a machine learning system that classifies half-court possessions into zone and man-to-man defensive categories.

Rather than relying only on the location of an individual defender, the model examines how the defensive unit responds to offensive movement throughout the possession.

Potential applications include identifying tactical trends and evaluating offensive effectiveness against different defensive structures.

Intro to Handle Score | NBA, October 20, 2026 Official description of the model, training data, and underlying player tracking.

Detecting Zone Defense and Measuring Its Impact | NBA, October 20, 2026 Explanation of possession-level defensive classification.

Baseball Analytics

Baseball has a long history of quantitative analysis, making it an important testing ground for increasingly sophisticated AI systems.

Modern baseball analytics combines statistical modeling, ball tracking, bat tracking, biomechanics, and automated video analysis.

Statcast and Player Tracking

MLB's Statcast system measures and analyzes characteristics of pitching, hitting, fielding, and baserunning.

Applications include:

  • Pitch velocity and movement.
  • Pitch release characteristics.
  • Exit velocity.
  • Launch angle.
  • Estimated hit outcomes.
  • Defensive positioning.
  • Sprint speed.
  • Bat speed.
  • Swing path.
  • Batted-ball trajectories.

These measurements help separate an observed result from characteristics of the underlying action.

For example, a hard-hit ball may be caught, while a poorly struck ball may result in a hit.

Expected-outcome models can help explain such differences probabilistically.

However, an expected outcome is not guaranteed, and the usefulness of any statistic depends on the question being investigated.

Bat Tracking and Swing Mechanics

MLB expanded its publicly described swing-analysis measurements during 2024 and 2025.

Newer bat-tracking metrics include:

  • Bat speed: A measurement of bat velocity during the swing.
  • Swing length: The distance traveled by the bat over a defined part of the swing.
  • Attack angle: The bat's vertical travel angle near the point of contact.
  • Attack direction: The bat's horizontal movement direction.
  • Swing path: The geometry of the bat's movement.
  • Contact location: Where the bat and ball interact.

These variables provide more direct information about swing mechanics than a batting average alone.

A particular attack angle is not universally optimal. Its effectiveness depends on pitch location, timing, bat orientation, and the hitter's objectives.

The MLB Statcast team discussed the development of these measurements at the 2025 Society for American Baseball Research (SABR) Analytics Conference.

New Statcast Metrics Measure Swing Path, Attack Angle, and Attack Direction | MLB, 2025 Explanation of the newer swing-analysis measurements.

MLB Statcast Updates: Baseball Savant, ABS, and Game Notes | SABR, 2025 Discussion by MLB analysts of the development and interpretation of tracking metrics.

Automated Ball-Strike Challenge System

Major League Baseball introduced the Automated Ball-Strike (ABS) Challenge System during the 2026 MLB season.

The system uses Hawk-Eye tracking technology to evaluate challenged ball-and-strike calls.

Unlike a fully automated strike-calling system, the 2026 challenge format retains a human home-plate umpire making the initial decision.

Eligible players can challenge a call, after which the tracking system determines whether the pitch crossed the defined automated strike zone.

The ABS zone uses measurements tied to each hitter's height.

According to MLB:

  • The zone is 17 inches wide.
  • Its upper limit is 53.5% of the hitter's measured height.
  • Its lower limit is 27% of the hitter's measured height.
  • Ball position is evaluated at the middle of home plate.

These specifications define the automated system's operational strike zone and should not be assumed to describe every historical interpretation of the strike zone.

The first regular-season MLB ABS challenge occurred on March 25, 2026, during the Yankees-Giants game.

The introduction of ABS illustrates a broader trend: machine-based geometric measurements are increasingly used to support, review, or standardize officiating decisions.

MLB Announces ABS Challenge System for 2026 | MLB, September 2025 Official adoption announcement and testing history.

ABS Challenge Dashboard | MLB Baseball Savant Technical rules and challenge data.

Automated Ball-Strike Challenge System Firsts | MLB, September 2026 Historical record of the system's initial MLB use.

Baseball Biomechanics

Pitching and hitting involve coordinated motion across multiple joints and body segments.

AI can assist in estimating:

  • Shoulder and elbow motion.
  • Trunk rotation.
  • Pelvic rotation.
  • Arm-slot consistency.
  • Stride mechanics.
  • Bat and hand trajectories.
  • Release-point variability.
  • Timing of the kinetic chain.

Research often combines motion capture with force measurements and biomechanical models.

A video-derived joint angle is a geometric measurement. It should not automatically be interpreted as a direct measurement of tendon force, ligament stress, or injury probability.

The American Baseball Biomechanics Society and SABR provide research forums where these measurement challenges are discussed.

SABR Analytics Conference Videos, 2025 | SABR Includes discussions of biomechanics, Statcast, computer vision, and quantitative baseball research.

Baseball - Stealing Signs

Machine learning can recognize patterns in visible baseball signals, including sequences of gestures used to communicate pitch selection.

In a widely circulated 2019 demonstration, engineer and science communicator Mark Rober explored how a phone-based machine learning system could identify patterns in baseball signs.

The demonstration illustrates a broader principle: repeated visual signals can be analyzed computationally when examples are available.

This application also raises questions about competitive integrity and the unauthorized use of technology.

Sports organizations distinguish between permitted observation and prohibited electronic sign-stealing according to their applicable rules.

Modern encrypted or restricted electronic communication systems have changed how some teams communicate pitch selections, reducing reliance on visibly transmitted signals.

Historical video: Stealing Baseball Signs with a Phone (Machine Learning)

Load video
YouTube

Mark Rober's 2019 demonstration explores machine learning applied to baseball signal recognition. Although shorter than a typical technical lecture, it remains a useful historical example of computer vision, pattern recognition, and the security implications of sports analytics.

Artificial Intelligence, Machine Learning, and the Bright Future of Baseball | Society for American Baseball Research Discussion of AI applications and the potential implications of machine-assisted sign recognition.

Skeletal Motion

Skeletal motion analysis uses computer vision to estimate the positions of anatomical landmarks and reconstruct human movement.

It is especially important in sports where technique, joint configuration, balance, and rotational motion affect evaluation or performance.

Two-Dimensional and Three-Dimensional Pose Estimation

A 2D pose estimator identifies joints in image coordinates.

A 3D pose estimator attempts to infer joint positions in three-dimensional space.

Possible approaches include:

  • Multi-camera triangulation.
  • Depth sensing.
  • Learned 3D pose estimation from ordinary video.
  • Marker-based motion capture.
  • Markerless motion capture.
  • Combining optical measurements with IMUs.

Multi-camera systems can reduce ambiguities that arise when reconstructing three-dimensional movement from a single camera.

However, they require calibration and synchronization.

Markerless systems can simplify data collection, but performance varies with lighting, camera placement, occlusion, clothing, body position, and the movements being analyzed.

Fujitsu AI Gymnastics Judging

Fujitsu and the International Gymnastics Federation, now branded as World Gymnastics, began collaborating in 2017 to develop AI-assisted judging technology.

Early research used advanced sensing and deep learning to estimate gymnast skeletal movements and identify performance elements.

Historical technical descriptions discussed multi-view depth images, three-dimensional joint reconstruction, and a movement database containing hundreds of recognized gymnastics elements.

An early account described more than 800 men's elements and more than 500 women's elements in the development database. Those numbers reflect an earlier stage of the project and should not be treated as the current system's complete coverage.

Meet Fujitsu's AI Gymnastics Judges | Synced Historical discussion of the early technology and its research challenges.

Judging Support System

The Fujitsu Judging Support System (JSS) has progressed beyond its early experimental design.

Fujitsu and the International Gymnastics Federation reported deployment across all ten artistic gymnastics apparatuses at the 2023 World Championships in Antwerp.

The system moved toward camera-based image analysis rather than relying on the earlier sensor configuration.

Its functions include:

  • Recognizing gymnastics elements.
  • Providing numerical motion information.
  • Supporting multi-angle video review.
  • Helping judges examine difficult skills.
  • Improving consistency in applying judging criteria.

World Gymnastics documents further trials in 2025, including expanded use by supervisors on selected apparatuses.

Its published deployment roadmap identifies further expansion of supervisory access during the 2026 World Championships.

The role of human judges remains essential.

An automated system can help identify measurable components of a performance, but official judging also involves interpreting rules, reviewing context, and making decisions that may not be fully determined by skeletal geometry.

Judging Support System | World Gymnastics Official history, competition deployments, and current functions.

Fujitsu and FIG Introduce AI Judging Support Across Ten Apparatuses | Fujitsu, October 2023 Description of the camera-based system and competition implementation.

Research Challenges in Skeletal Motion

Current technical challenges include:

  • Reliable pose estimation during fast rotations.
  • Tracking athletes through occlusion.
  • Reconstructing movement when body parts overlap.
  • Estimating depth from limited camera viewpoints.
  • Distinguishing similar skills.
  • Accounting for anatomical differences.
  • Calibrating measurements against reference systems.
  • Quantifying uncertainty in joint positions and angles.

A system may recognize an athletic movement accurately in a controlled environment while performing less reliably during live competition.

Validation should include different athletes, apparatuses, venues, movement types, and camera configurations.

Equipment Design and Sports Engineering

AI is also used in designing and evaluating sporting equipment.

Optimization methods can search large spaces of possible equipment configurations, while simulation models estimate mechanical performance.

Applications include:

  • Golf club geometry and mass distribution.
  • Baseball bat design.
  • Footwear cushioning and stability.
  • Helmet impact protection.
  • Bicycle aerodynamics.
  • Prosthetic sporting equipment.
  • Sports-surface analysis.

AI in Golf Club Design

AI-assisted golf club design is an established example of computational optimization.

Machine learning and simulation can help engineers investigate the relationship between club geometry, mass distribution, impact conditions, and ball flight.

These methods can reduce the number of physical prototypes required to explore potential designs.

However, the predicted mechanical performance of a design still requires physical testing.

AI Is Changing the Face of Golf Club Design | Steve McCaskill, Forbes, January 2019 Historical example of AI-assisted sports equipment engineering.

Digital Twins and Simulation

A digital twin is a computational representation of a physical object or system that can be updated using measurement data.

Possible sports applications include:

  • Modeling athletic equipment under different loads.
  • Investigating alternative playing-surface conditions.
  • Simulating ball trajectories.
  • Analyzing motion sequences.
  • Exploring changes in movement mechanics.

A validated physics-based simulation can help investigate design choices.

A generative simulation that looks realistic does not necessarily reproduce real physical behavior.

Models should be tested against instrumented experiments, observed trajectories, and known mechanical constraints.

Fan Engagement and Viewing Experience

AI can make detailed sports information easier to access and understand.

Automated Highlights and Content Discovery

Machine learning can identify potential highlights from video, audio, event data, and audience signals.

Common applications include:

  • Detecting goals, shots, and scoring plays.
  • Identifying important game moments.
  • Generating searchable video collections.
  • Producing personalized highlight packages.
  • Creating automated subtitles and descriptions.
  • Summarizing matches.

Advanced systems may combine event recognition with language generation to produce explanations of plays.

Such systems must distinguish documented events from inferred interpretations.

Generated commentary can contain factual errors, particularly when player identity, event timing, or game context is ambiguous.

Immersive and Interactive Viewing

AI can support virtual and augmented reality applications.

Examples include:

  • Reconstructed three-dimensional plays.
  • Virtual camera perspectives.
  • Player trajectory overlays.
  • Tactical diagrams.
  • Interactive statistical graphics.
  • Simulation-based educational experiences.

A reconstructed camera angle should be identified as a visualization rather than mistaken for original camera footage.

AI Assistants for Fans and Analysts

Natural-language interfaces can allow users to ask questions such as:

  • Which players created the most passing opportunities?
  • How did defensive positioning change after a substitution?
  • Which pitches produced the highest swing-and-miss rates?
  • How frequently did a team use zone defense?
  • Which athletes showed the largest increase in sprint distance?

For reliable answers, an AI assistant should retrieve information from validated statistics and tracking databases.

A general-purpose language model without access to reliable data may generate plausible but incorrect sports statistics.

Why This Matters

The growing use of AI in sports has implications beyond improving team performance.

Scientific Importance

Sports provide a demanding research environment for understanding human motion and multi-agent interaction.

Sporting events involve:

  • Multiple interacting decision-makers.
  • Rapid changes in physical conditions.
  • Incomplete information.
  • Complex spatial relationships.
  • Measurable outcomes.
  • High-speed physical movement.

These characteristics make sports useful for developing and evaluating computer vision, temporal modeling, reinforcement learning, and multi-agent systems.

Progress in sports analytics can inform other research areas involving movement and coordination, although successful transfer is not automatic.

Coaching and Competitive Performance

AI can reduce the manual effort required to review large quantities of video and tracking data.

It can reveal patterns that are difficult to identify consistently through observation alone.

Potential benefits include:

  • More systematic opponent preparation.
  • Improved training feedback.
  • Faster retrieval of comparable situations.
  • Better evaluation of tactical alternatives.
  • More detailed athlete development records.

The value of these benefits depends on whether information actually improves decisions.

Officiating and Consistency

Instrumented measurements can help officials evaluate events involving precise position or timing.

Examples include football offside assessments, tennis line calling, and baseball strike-zone challenges.

However, sports rules vary in how much judgment they require.

A measurement system can provide precise coordinates while still depending on human interpretation of a rule.

Good implementation requires documented procedures for equipment failure, disputed measurements, and final decision authority.

Accessibility and Competitive Equity

Historically, sophisticated sports analytics required expensive data infrastructure and specialized personnel.

Improved software, shared research datasets, and cloud computing can reduce some barriers.

However, costs remain associated with cameras, sensor systems, data licensing, expertise, and ongoing validation.

Wider access to analytical tools does not guarantee equal competitive resources.

Athlete Welfare and Privacy

Athletic performance data can reveal sensitive information about health, physical capabilities, and employment-related risks.

Important governance questions include:

  • Who owns the tracking and biometric data?
  • Can athletes access or correct their records?
  • How long is the data retained?
  • Can measurements be used in contract negotiations?
  • Who may access injury-risk estimates?
  • What happens when a model makes a harmful prediction?
  • Can an athlete challenge an automated assessment?

These questions should be addressed through clear agreements, appropriate safeguards, and human oversight.

Research and Development

Recent Developments

The period from 2024 through October 2026 has brought significant advances in sports tracking, computer vision, computational tactics, and AI-supported officiating.

Some developments are already operational in professional competition. Others are research demonstrations whose practical impact remains unverified.

2026: Advanced Offside Technology and FIFA Football AI Pro

During the 2026 FIFA World Cup, FIFA deployed an advanced version of semi-automated offside technology.

The system was designed to provide quicker alerts for clear positional offside situations while retaining human decision-making for more complex interpretations.

FIFA also incorporated detailed skeletal tracking and three-dimensional player representations.

A major additional development was Football AI Pro, a generative AI knowledge assistant intended to make advanced match analysis accessible to all 48 participating national teams.

The assistant helps users retrieve and interpret detailed performance information, rather than requiring every coaching staff to manually process extensive technical reports.

FIFA also described generating more than 150 million tracking data points per match from its stadium infrastructure.

What was demonstrated:

  • Operational deployment of advanced tracking-based officiating.
  • Automated assistance for clear positional offside decisions.
  • Three-dimensional reconstruction for officiating and visualization.
  • Broader access to match-analysis tools.

What remains uncertain:

  • How much AI-assisted analysis improves team decisions or results.
  • Whether comparable benefits can be delivered at lower infrastructure cost.
  • How reliably reconstructed match situations capture player intentions.
  • How results generalize across different competitions and camera installations.

Faster Offside Decisions and Football AI Pro | FIFA, June 3, 2026 Official description of the technologies introduced at the 2026 World Cup.

2026: Player Tracking and Group Activity Recognition

A 2026 CVPR Workshops paper, Pixels or Positions? Benchmarking Modalities in Group Activity Recognition, investigated whether team activities are better recognized from video images or tracking coordinates.

The researchers introduced the SoccerNet-GAR benchmark, using synchronized video, tracking, and event information from the 2022 FIFA World Cup.

The dataset included 87,939 annotated group activities across ten categories.

The paper compared video-based recognition with a graph model using player positions and tactical relationships.

Its tracking-based baseline achieved 70.4% balanced accuracy and 54.6% macro F1, compared with 56.8% balanced accuracy and 46.8% macro F1 for the strongest evaluated video baseline.

The tracking model also required fewer parameters and less training computation in the reported comparison.

Interpretation: Structured positional information can be highly effective for recognizing collective sporting actions.

Limitation: These findings apply to the defined benchmark and tested methods. They do not show that player coordinates are always superior to video or that the same results will generalize to every sport.

Pixels or Positions? Benchmarking Modalities in Group Activity Recognition | Karki et al., CVPR Workshops, 2026 Research on multimodal sports datasets and graph-based group activity recognition.

2026: Improved Multi-Object Tracking

The September 2026 research paper SportsTrack: A Dual-Stream Prediction and Multi-Stage Association Tracking Framework for Sports Video Analysis examined the problem of following athletes through complex motion and visual occlusion.

Its framework combines motion prediction and association methods, including recurrent neural-network components and Kalman-based modeling.

The researchers reported improvements on the SportsMOT benchmark using established tracking metrics.

These include HOTA, which evaluates detection and association quality, and IDF1, which measures identity consistency.

Such methods address a central problem in sports analytics: maintaining the correct identity of a player through a fast, crowded sequence.

Limitation: Benchmark improvements do not automatically establish performance across every sport, camera system, or competition setting.

SportsTrack: A Dual-Stream Prediction and Multi-Stage Association Tracking Framework for Sports Video Analysis | Yang et al., Image and Vision Computing, 2026 Research on nonlinear movement prediction and multi-object association.

2026: NBA Preference-Based Skill Metrics

In October 2026, the NBA published its new Handle Score methodology.

The system combines high-frequency skeletal tracking with expert pairwise evaluations of basketball plays.

This is an important research direction because it attempts to model a complex sporting skill for which no simple physical measurement provides a complete answer.

The use of preference learning introduces a methodological question: whose judgments define the target?

Different expert groups may disagree about the importance of creativity, efficiency, difficulty, or game context.

Independent evaluation should examine whether the score predicts meaningful performance outcomes and whether it remains stable across playing styles.

2026: Injury Prediction Research Identifies Persistent Validation Problems

Systematic reviews published in 2026 reinforced concerns about sports injury prediction models.

Many studies report promising results but use relatively small or restricted datasets.

External validation, calibration, consistent injury definitions, and prospective testing remain major weaknesses.

This development is important because the field is moving from asking whether machine learning can identify patterns to asking whether those patterns are sufficiently trustworthy for clinical decisions.

The evidence does not yet justify treating an injury prediction model as a substitute for medical assessment.

2025: Advances in Football Video Analytics

A July 2025 study in Scientific Reports investigated real-time ball-player interaction analysis.

The researchers combined computer vision detection with graph convolutional networks to model player relationships, distances, movements, and event information.

The study reported favorable results across several benchmark measurements.

The findings support the feasibility of integrating object detection, tracking, and relational analysis into a unified processing system.

However, the reported performance reflects specific evaluation datasets and metrics. Deployment in new competition environments requires further validation.

Real-Time Analysis of Soccer Ball-Player Interactions Using Graph Convolutional Networks | Majeed et al., Scientific Reports, July 2025 Peer-reviewed study of detection, tracking, and graph-based analysis.

2025: Reconstructing Missing Tracking Data

Broadcast footage often fails to show all players at once.

A 2025 research effort investigated reconstructing football tracking data using a multimodal transformer and diffusion framework.

The approach combined incomplete visual information with event data and generative modeling.

Its objective was to produce plausible, temporally consistent trajectories even when parts of the original scene were not visible.

The researchers compared reconstructed positions with in-venue tracking measurements.

This approach could increase the usefulness of historical broadcast archives that lack dedicated tracking systems.

Important qualification: Reconstructed trajectories are model estimates, not direct observations. They should be accompanied by uncertainty information and should not be used as definitive evidence of an unseen player's exact position.

Multimodal Transformer-Diffusion Framework for Large-Scale Reconstruction of Soccer Tracking Data | Hughes et al., 2025 Research on recovering football trajectories from incomplete visual observations.

2025-2026: Machine Learning in Professional Football and Baseball

The NFL's 2025 defensive coverage models illustrate an increasing emphasis on recognizing tactical responsibilities rather than merely measuring player movement.

MLB's 2026 ABS deployment demonstrates how tracking data can become part of an operational officiating process.

These developments represent different applications of AI:

  • Tactical inference: Estimating roles and responsibilities that are not directly visible.
  • Measurement-based review: Applying a defined rule to measured positions.
  • Decision support: Presenting outputs to a human decision-maker.
  • Operational validation: Testing whether a system performs reliably during actual competition.

The distinction matters because these systems require different validation procedures.

Current Academic Debates

Prediction Versus Causation

Machine learning models often discover statistical relationships without identifying the mechanisms producing them.

For example, a model may find an association between high training load and injury incidence.

That association does not establish that reducing load by a particular amount will prevent injury.

Other variables, including competition intensity, fatigue, previous injury, and training adaptation, may influence both workload and injury outcomes.

Causal claims require appropriate study designs and careful consideration of confounding variables.

Generalization Across Teams and Leagues

A model trained on one league may perform poorly elsewhere.

Potential reasons include:

  • Different tactical styles.
  • Different athlete populations.
  • Different rules.
  • Different tracking hardware.
  • Different playing surfaces.
  • Different labeling procedures.
  • Changes in competition intensity.

A model should be evaluated using data that genuinely differs from its development sample.

Data Leakage

Sports datasets contain repeated observations of the same athletes, teams, and competitions.

Randomly splitting individual frames or observations between training and testing can produce misleadingly favorable results.

For example, nearly identical frames from the same play may appear in both datasets.

Appropriate evaluation may require separation by match, athlete, team, season, venue, or chronological period.

The correct approach depends on the intended deployment.

Interpretability

Coaches and clinicians need to understand why a model recommends an action.

Attention maps, feature importance, and post-hoc explanations can help, but these tools do not automatically provide causal explanations.

An explanation should be tested for stability and usefulness, not merely for whether it sounds plausible.

Human-AI Collaboration

Experts may benefit from AI-generated suggestions, but they may also become excessively reliant on automated outputs.

Systems should present uncertainty and allow users to challenge incorrect interpretations.

Human oversight must be meaningful, particularly in medical, officiating, and athlete-selection decisions.

Practical Implementation Guide

A sports organization introducing AI should begin with a measurable problem rather than purchasing technology without a clear objective.

Example Development Workflow

  1. Define the decision. Identify what the model will support, such as detecting changes in sprint performance or identifying passing options.
  2. Identify data sources. Determine which measurements are available, permitted, and sufficiently reliable.
  3. Establish ground truth. Create trustworthy reference labels or instrumented measurements.
  4. Build a baseline. Compare simple statistical methods before adopting complex neural networks.
  5. Train and validate. Use appropriate temporal, player-level, or competition-level data separation.
  6. Measure uncertainty. Report errors and confidence or calibration information where appropriate.
  7. Evaluate operational constraints. Test latency, equipment failure, camera differences, and missing data.
  8. Involve domain experts. Ask coaches, athletes, officials, or clinicians to assess practical usefulness.
  9. Conduct prospective testing. Determine whether model-supported decisions improve the intended outcome.
  10. Monitor after deployment. Check for data drift, changing athlete behavior, and unexpected consequences.

Evaluation Metrics

Task Common metrics Important limitation
Object detection Precision, recall, mean average precision A correct bounding box does not guarantee correct player identity
Multi-object tracking HOTA, IDF1, identity switches A model may track visible players but miss off-screen participants
Pose estimation Joint position error, percentage of correct keypoints Pose accuracy does not automatically establish kinetic accuracy
Action recognition F1-score, balanced accuracy, event-level precision and recall Class imbalance and ambiguous event boundaries affect results
Probability prediction Log loss, Brier score, calibration, discrimination Good predictions do not establish causal relationships
Injury risk prediction Sensitivity, specificity, precision, calibration, external validation Rare events can make conventional accuracy misleading
Tactical recommendations Expert assessment, simulations, prospective outcome measures Expert preference is not equivalent to demonstrated competitive improvement
Officiating systems Measurement error, operational latency, reliability, auditability Accurate geometry may not resolve subjective rule interpretation

Questions to Ask Technology Vendors

  • Was the system independently validated?
  • Was it evaluated using data from the same sport, competitive level, and camera environment?
  • Are measurement errors and uncertainty reported?
  • What reference standard was used?
  • Can the results be reproduced?
  • How does the system perform for athletes outside its training population?
  • What happens when sensors fail or video quality deteriorates?
  • Can users inspect or contest the output?
  • Who controls the collected data?
  • Is there evidence that the system improves real-world decisions?

These questions are especially important when a product claims to prevent injuries, identify future elite athletes, or produce optimal tactical decisions.

Open Questions and Future Research

Several important problems remain unresolved.

1. Can sports foundation models generalize across leagues and competitions?

Large pretrained models may reduce the amount of labeled data needed for individual tasks, but differences in sports rules, camera configurations, and athlete behavior complicate transfer.

2. Can AI reliably model collective intentions?

A tracking system can measure movement but may not know whether a player intentionally created space, followed instructions, or reacted to an opponent.

Inferring intent from movement alone remains difficult.

3. How should uncertainty be communicated?

Single-number scores can conceal uncertainty caused by missing information, labeling disagreements, and changing game conditions.

More useful systems may provide ranges, alternative explanations, and sensitivity analyses.

4. Can simulation-based coaching recommendations be validated prospectively?

Generative models can propose realistic tactics, but the important question is whether those tactics improve real performance when adopted.

5. Will biomechanical AI improve injury prevention?

Reliable evidence will require prospective studies, appropriate reference measurements, diverse athlete populations, and demonstrated improvements in health outcomes.

6. How should automated officiating balance speed, precision, and fairness?

A more precise measurement can still create controversy when rules require interpretation or when measurement uncertainty is poorly communicated.

7. Can advanced sports analytics become accessible outside elite professional organizations?

Lower-cost cameras, open research software, and shared datasets may broaden access, but robust deployment still requires calibration, expertise, data protection, and maintenance.

Additional Viewing

The following presentations offer technical or research-oriented perspectives on AI in sports.

AI Football Tactics: Petar Veličković

Load video
YouTube

Petar Veličković of Google DeepMind discusses TacticAI in a presentation associated with Data Science Conference Europe 2023, published on YouTube in April 2024. The approximately 34-minute talk explains the challenge of modeling football tactics and the use of AI as a decision-support system developed with Liverpool FC.

Watch on YouTube

AI in Emerging Sports Leagues: MIT Sloan, 2026

Load video
YouTube

This approximately 45-minute MIT Sloan Sports Analytics Conference panel examines applications of AI in emerging sports leagues, including officiating, operations, athlete health, and fan engagement. Participants discuss implementation challenges and the need to balance technological opportunities with practical league requirements.

Watch on YouTube

MLB Statcast: Baseball Savant, ABS, and Game Notes

Load video
YouTube

This 2025 SABR Analytics Conference presentation features members of MLB's Statcast team discussing tracking technology, Baseball Savant, automated ball-strike systems, and analytical reporting. It is particularly useful for understanding how raw measurements become practical statistics and decision-support resources.

Watch on YouTube

Baseball Biomechanics and Big Data

Load video
YouTube

This SABR 2025 panel, presented with the American Baseball Biomechanics Society, brings together specialists associated with professional baseball organizations and biomechanics research. The discussion connects motion analysis and quantitative measurements with the practical challenges of applying biomechanical information to player development.

Watch on YouTube

MLB Statcast Updates, 2026

Load video
YouTube

This 2026 SABR Analytics Conference presentation features MLB representatives discussing developments in Statcast and baseball measurement technology. It provides a more recent companion to the 2025 presentation and illustrates the continuing evolution of baseball analytics.

Watch on YouTube

Original Sports AI Video Collection

The original page contained an extensive collection of embedded videos spanning sports AI, baseball analytics, and skeletal motion. The following legacy links are retained as a historical index rather than presenting unverified video descriptions or assuming that every older upload remains available.

General sports AI and motion analysis

Original baseball analytics video collection

Original skeletal motion and gymnastics videos

Research Papers and Technical References

Peer-Reviewed Research

Datasets and Research Platforms

Sports Organizations and Technology Documentation

Further Reading and Historical Coverage

The following sources provide useful historical context about the development of AI applications in sports.

Conclusion

Artificial intelligence has moved from an experimental tool in sports research to an important component of professional performance analysis, competition operations, and broadcast production.

The strongest established applications involve measuring observable events, processing large tracking datasets, recognizing movement patterns, and supporting decisions where relevant data can be validated.

More ambitious applications, including personalized injury prevention, fully automated tactical optimization, and generative simulations of future athletic performance, remain active research areas.

Recent advances in computer vision, graph neural networks, transformers, sensor systems, and foundation-model research provide powerful new methods for studying athletic movement and team behavior.

The central challenge is no longer simply collecting more data. It is determining which measurements are trustworthy, which predictions generalize, which interventions improve outcomes, and how these systems can be integrated responsibly into sports.

The future of sports AI will depend on combining rigorous technical validation with coaching expertise, athlete welfare, transparent officiating, and meaningful human control.

Retrieved from "https://primo.ai/w/index.php?title=Sports&oldid=49324"