FST.ai for Taekwondo Officiating

• FST.ai is an AI-assisted officiating framework that leverages real-time video capture, pose estimation, and action recognition to support rapid head kick scoring in Taekwondo.
• FST.ai 2.0 expands the system into an explainable AI ecosystem with interactive dashboards, uncertainty modeling, and multi-stakeholder tools for referee training and fairness monitoring.
• The system demonstrates significant operational improvements by reducing review times from up to 90 seconds to as little as 4.6 seconds while enhancing referee trust with clear, audit-ready overlays.

Searching arXiv for papers explicitly about “FST.ai” and closely related variants to ground the article in current literature. arXiv search query: "FST.ai" FST.ai is an AI-assisted officiating framework for Sport Taekwondo centered on fair, fast, and transparent decision support in high-speed scoring situations, especially real-time head kick detection and the distinction between standard and turning head kicks (Shariatmadar et al., 19 Jul 2025). In later work, the concept is expanded into “FST.ai 2.0,” an explainable AI ecosystem for Olympic and Paralympic Taekwondo that combines pose-based action recognition, uncertainty-aware inference, explainability overlays, interactive dashboards, referee education tools, fairness monitoring, and Para-Taekwondo classification assistance (Shariatmadar et al., 21 Oct 2025). In current arXiv usage, “FST.ai” most directly denotes this sports-officiating line of work rather than other unrelated uses of the initials “FST” in machine learning, finite-state transducers, or neuroimaging (Shariatmadar et al., 19 Jul 2025, Shariatmadar et al., 21 Oct 2025).

1. Definition, scope, and problem setting

FST.ai is introduced as a response to the shortcomings of traditional manual officiating and Instant Video Replay (IVR) in Taekwondo. The stated failure modes are latency, subjectivity, inconsistency, operational disruption, and erosion of athlete and coach trust. The initial paper frames the core use case as real-time head kick detection and scoring, where scoring events occur in fractions of a second and the distinction between a valid head kick and a turning head kick has direct point consequences (Shariatmadar et al., 19 Jul 2025).

The system is explicitly described as a decision-support framework rather than a referee replacement system. Its intended role is to provide a recommendation to the review jury within seconds while keeping the human jury as the final authority. In the later ecosystem formulation, this role broadens beyond scoring support to include referee evaluation, athlete performance analysis, Para-Taekwondo classification, and governance-oriented fairness analytics (Shariatmadar et al., 21 Oct 2025).

The 2.0 formulation makes the scope multi-stakeholder. Referees and review juries use it for live decision support and post-match analysis; coaches and athletes use it for tactical and performance review; Para-Taekwondo classifiers use it for uncertainty-aware motor assessment; and governing bodies use it for consistency monitoring, fairness auditing, and policy analytics. This suggests a shift from a single officiating module toward an integrated competition-and-training infrastructure (Shariatmadar et al., 21 Oct 2025).

2. Development and naming

The first explicit arXiv paper titled “FST.ai” defines the acronym as “Fairness, Speed, and Trust in AI officiating” and presents a modular pipeline for Taekwondo officiating under the R3AL.ai project, with head kick scoring as the flagship scenario (Shariatmadar et al., 19 Jul 2025). The later paper, “FST.ai 2.0,” retains the fairness and speed emphasis but redefines the system as an explainable AI ecosystem for fair, fast, and inclusive decision-making in Olympic and Paralympic Taekwondo (Shariatmadar et al., 21 Oct 2025).

The progression from the first paper to 2.0 is not merely incremental model refinement. The original system centers on live video capture, pose estimation, temporal action recognition, impact verification, score suggestion, human confirmation, and logging. The later system adds graph convolutional action recognition, epistemic uncertainty modeling through credal sets, explainability overlays, interactive dashboards, referee education modules, fairness monitoring, and Para-Taekwondo classification support (Shariatmadar et al., 19 Jul 2025, Shariatmadar et al., 21 Oct 2025).

The term is also potentially name-confusable in the broader arXiv ecosystem. Other papers use closely related acronyms for unrelated topics, including few-shot teamwork, functional spatiotemporal Mamba, finite-state transducers, and “First Shape, Then Meaning.” A plausible implication is that “FST.ai” requires domain context for correct interpretation. In present explicit arXiv titling, however, the direct match is the Taekwondo officiating framework and its 2.0 extension (Fosong et al., 2022, Wei et al., 2024, Ginn et al., 16 Jan 2026, Chierchia et al., 5 May 2026, Shariatmadar et al., 19 Jul 2025, Shariatmadar et al., 21 Oct 2025).

3. Core architecture and computational workflow

The original FST.ai pipeline consists of live video capture, preprocessing, pose estimation, temporal action recognition, impact/contact verification, score suggestion, human-in-the-loop confirmation, and score registration with logging (Shariatmadar et al., 19 Jul 2025). The primary input is high-speed video from strategically positioned cameras around the mat, with frame rates exceeding 60 fps and both wide-angle and athlete-centered views. The video stream is denoted

$I = \{I_t\}_{t=1}^{T}.$

Preprocessing is described as

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$

The paper also gives a deblurring relation in corrupted form, intended as a deconvolution of a blurred frame with an estimated blur kernel (Shariatmadar et al., 19 Jul 2025).

Pose estimation converts each frame into a skeleton:

$f_{pose}: I_t \rightarrow P_t,$

with pose output

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$

The initial paper states that FST.ai uses OpenPose or a similar network with confidence maps and Part Affinity Fields. It further proposes heuristic pre-flagging of possible head kicks based on ankle-head geometry, and temporal joint tracking via Kalman filters or optical flow alignment, with update equation

$$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$

This stage transforms raw video into structured kinematic trajectories suitable for action recognition (Shariatmadar et al., 19 Jul 2025).

Action recognition in the original paper operates on temporal sliding windows over pose sequences:

$$f_{class}: \mathbb{R}^{n\times 2K} \rightarrow \{A_1, A_2, \dots, A_m\}.$$

The paper mentions CNN-LSTM hybrids and Transformer encoders, together with features such as joint angles

$\theta_i = \angle (p_{i-1}, p_i, p_{i+1}),$

and velocities

$v_i = \frac{|p_i^{(t)} - p_i^{(t-1)}|}{\Delta t}.$

The main class set in the head-kick use case is slide/invalid motion, standard head kick, and turning head kick, corresponding to score suggestions of 0, 3, and 5 points (Shariatmadar et al., 19 Jul 2025).

FST.ai 2.0 replaces or supplements this sequence model with skeleton-based graph learning. It represents each pose sequence as a spatial-temporal graph $G=(V,E)$, with joint tensor

$\mathbf{X} \in \mathbb{R}^{T \times N \times C},$

and adjacency

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$0

A standard ST-GCN layer is given as

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$1

with prediction

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$2

The 2.0 paper also introduces a transformer-based spatiotemporal model over pose embeddings, with self-attention

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$3

and a loss of the form

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$4

This suggests a broadened perception stack rather than a single fixed model family (Shariatmadar et al., 21 Oct 2025).

Impact verification in the original paper is separate from classification. A scoring event requires both valid motion and verified contact, operationalized through foot deceleration and spatial overlap between foot and head regions. The intended rule is

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$5

Turning-kick qualification is then assessed from torso rotation, with the paper giving conceptually consistent but numerically varying examples above $$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$6, $$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$7, and $$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$8 (Shariatmadar et al., 19 Jul 2025).

4. Uncertainty, explainability, and human oversight

A defining feature of FST.ai 2.0 is that action prediction is coupled to explicit uncertainty modeling. The paper distinguishes aleatoric and epistemic uncertainty. Aleatoric uncertainty is expressed through outputs of the form

$$I_t' = \phi(I_t) = \text{normalize}(\text{deblur}(\text{hist\_eq}(I_t))).$$9

while epistemic uncertainty is estimated through approximate Bayesian procedures such as Monte Carlo dropout:

$f_{pose}: I_t \rightarrow P_t,$0

The uncertainty decomposition is written as

$f_{pose}: I_t \rightarrow P_t,$1

This supports abstention, escalation, and review rather than unconditional automation (Shariatmadar et al., 21 Oct 2025).

The 2.0 paper further adopts credal sets and interval-valued probabilities:

$f_{pose}: I_t \rightarrow P_t,$2

For Para-Taekwondo classification it defines

$f_{pose}: I_t \rightarrow P_t,$3

and also presents bounded class probabilities

$f_{pose}: I_t \rightarrow P_t,$4

A plausible implication is that the system is designed to preserve ambiguity when evidence is insufficient, rather than collapsing every case to a single hard label (Shariatmadar et al., 21 Oct 2025).

The same paper gives an explicit conservative decision rule for sensor-fused next-generation scoring. With fused impact score

$f_{pose}: I_t \rightarrow P_t,$5

interval bounds

$f_{pose}: I_t \rightarrow P_t,$6

and imprecise probability bounds

$f_{pose}: I_t \rightarrow P_t,$7

a point is awarded only if

$f_{pose}: I_t \rightarrow P_t,$8

This is the clearest formal expression of how uncertainty affects final decision support (Shariatmadar et al., 21 Oct 2025).

Explainability is implemented through Grad-CAM, attention maps, keypoint overlays, kick trajectories, and confidence color bands. The Grad-CAM relation is given as

$f_{pose}: I_t \rightarrow P_t,$9

with

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$0

For transformer models, attention maps are rendered over joints and time, and the paper describes overlays that highlight the kicking foot, opponent’s head region, decisive frames, and uncertainty level. The stated purpose is to make AI-supported calls legible to referees and juries rather than opaque (Shariatmadar et al., 21 Oct 2025).

Both versions of the system retain explicit human authority. The original paper defines

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$1

with feedback

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$2

while the 2.0 paper gives an override rule

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$3

It also specifies audit logging:

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$4

This architecture supports contestability, retraining, and governance review (Shariatmadar et al., 19 Jul 2025, Shariatmadar et al., 21 Oct 2025).

5. Empirical evidence, deployment, and reported performance

The strongest operational claim of the original FST.ai paper is speed. It contrasts AI-assisted review with conventional IVR delays of 30–90 seconds and describes a concrete championship example in which one head-kick IVR review took nearly 90 seconds. It also gives a cumulative operational-loss example:

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$5

For on-device inference it states

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$6

with target

$P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$7

and example timings of $P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$8 ms for pose estimation, $P_t = \{(x_i, y_i, C_i)\}_{i=1}^K.$9 ms for action classification, and $$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$0 ms for impact detection, totaling $$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$1 ms. It then distinguishes this model-compute latency from the practical officiating response time of 3–5 seconds from action occurrence to jury-facing recommendation (Shariatmadar et al., 19 Jul 2025).

The 2.0 paper provides the most concrete field evidence. Its pilot deployment at the 2025 World Taekwondo Cadet Championships in Fujairah is described as using 68 recorded matches across 14 weight categories, 27 certified international referees, 6 jury members, synchronized integration with Daedo PSS, dual 120 fps 1080p cameras, and local inference on NVIDIA RTX 4090 hardware. It also reports 156 IVR requests across four competition days on a single-court deployment operating in parallel with official IVR infrastructure (Shariatmadar et al., 21 Oct 2025).

Quantitatively, the 2.0 abstract reports an 85% reduction in decision review time and 93% referee trust in AI-assisted decisions. Detailed sections report a reduction from about 89.7 seconds to 4.6 seconds, corresponding to

$$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$2

The paper does not reconcile this discrepancy, so the literature presently contains both numbers (Shariatmadar et al., 21 Oct 2025).

Reported accuracy values vary by task and subsystem. The original paper gives example confidences such as $$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$3 for a turning head kick and score recommendation confidence around 88–89%, but does not provide a conventional benchmark table with accuracy, precision, recall, or F1 (Shariatmadar et al., 19 Jul 2025). The 2.0 paper reports multiple task-specific values, including 94.2% for valid head-kick detection, 89.6% for spinning kicks, 92.4% for a transformer-based action recognizer on 300 annotated match segments, 92.8% head-kick classification accuracy in field validation, and an overall pilot decision accuracy of 0.927 against jury consensus (Shariatmadar et al., 21 Oct 2025).

The system also reports usability and governance-related metrics. The 2.0 paper gives a trust score of

$$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$4

interpreted as 93% referee trust, alongside 87% of referees and 93% of coaches rating AI assistance as valuable or very valuable. It further reports a reduction in override rate from

$$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$5

equivalent to

$$\hat{x}_{t|t}^j = \hat{x}_{t|t-1}^j + K_t (z_t^j - H \hat{x}_{t|t-1}^j).$$6

Additional claims include a 35% drop in variance among jury decisions, a 9.1% decision-consistency gain compared with manual IVR review, and 83% overlap between model saliency and referee focus maps (Shariatmadar et al., 21 Oct 2025).

For Para-Taekwondo, the evidence is more preliminary. The 2.0 paper reports 87.3% accuracy on an expert-verified test set with a 12.5% ambiguity flag rate and 2.8 seconds average decision time per athlete after preprocessing in one section; elsewhere it reports 88.7% alignment with expert panels across 78 classified Para-athletes and 9.5% re-evaluation triggers; and in another place it notes a smaller-scale test with four Para-athletes that was not statistically conclusive. This suggests an active but less mature validation status than the officiating module (Shariatmadar et al., 21 Oct 2025).

6. Broader significance, generalization, and limitations

The original FST.ai paper explicitly argues that the framework is not limited to Taekwondo. It identifies pose estimation, motion classification, and impact analysis as the transferable core, and names judo, karate, fencing, football, and basketball as plausible adaptation targets for foul recognition, action detection, or performance tracking (Shariatmadar et al., 19 Jul 2025). The 2.0 paper broadens this by emphasizing governance-aware, human-aligned AI in sport, with modules for referee education, fairness monitoring, policy analytics, and inclusive assessment (Shariatmadar et al., 21 Oct 2025).

At the same time, the literature describes clear limitations. The 2025 FST.ai paper is methodologically detailed at the pipeline level but thin in conventional ML validation: it does not provide dataset size, train/validation/test splits, optimizer settings, full benchmark tables, ablations, or standard metrics such as precision, recall, F1, ROC/AUC, or calibration error. It also contains threshold inconsistencies in turning-kick logic and acknowledges hard cases such as occlusion, minimal-touch contact, and deployment complexity (Shariatmadar et al., 19 Jul 2025).

The 2.0 paper is stronger empirically but still pilot-oriented. It provides descriptive metrics rather than comprehensive inferential statistics, does not report full confidence intervals or p-values for most claims, and presents subsystem accuracies drawn from different setups rather than a single harmonized benchmark. Data scarcity, especially for Para-Taekwondo, hardware constraints for real-time deployment, and stakeholder acceptance are explicitly noted as continuing challenges (Shariatmadar et al., 21 Oct 2025).

A balanced interpretation is therefore that FST.ai is best understood as an emerging class of explainable, uncertainty-aware sports-officiating systems rather than a finished autonomous judging platform. Its most substantiated capability is live referee support for Taekwondo review scenarios, especially head-kick scoring. Its broader ecosystem ambitions—Para-classification, federation-wide governance analytics, and generalized sport transfer—are technically plausible within the papers’ architectural framing, but remain less conclusively validated than the core officiating use case (Shariatmadar et al., 19 Jul 2025, Shariatmadar et al., 21 Oct 2025).