The Achilles heel of Poomsae: how can something designed to be subjective be judged objectively?

World Taekwondo’s Poomsae is based on a format that combines extreme subjectivity, cognitive overload, and inadequate tools for referees. The result: inconsistent scores, suspicions of bias, and growing discredit. The solution already exists and is ready for implementation: multi-camera computer vision and AI-assisted analysis, with pre-standardized metrics for angles, rhythm, and coordination.

1) The current format increases the disparity: “seeing two at the same time” is asking for the impossible.

Many tournaments use a face-to-face format (Chung vs. Hong) where two athletes perform simultaneously and the judges must score both in real time. This format, designed to speed up the competition and allow for “direct comparison,” multiplies the attentional load and encourages inconsistency among judges. The regulations of federations that adopt WT guidelines explicitly describe joint entry and simultaneous execution.

In this context, it is not surprising to see differences of up to several tenths between judges for recognized Poomsae, in Accuracy (4.0) and Presentation (6.0), categories that—by design—depend on assessments of rhythm, power, or “expression of energy.” Even with averages and the elimination of highs and lows, inter-judge variability persists because the input remains subjective.

“It is practically impossible to follow two athletes and score accurately on a touchscreen tablet; many watch the entire poomsae and enter the scores at the end from memory.”
— Testimonials from referees consulted by this media outlet, under reserve

2) Tools that do not help: touchscreen tablets without physical buttons

In Poomsae, touchscreen interfaces (Android tablets) are used to enter scores. Unlike buttons with physical feedback that exist in other contexts of electronic devices (e.g., judge boxes and triggers in sparring systems), the screen does not offer tactile differentiation and increases the risk of simultaneous errors. Approved suppliers for Poomsae promote Android tablets as the standard judging interface.

The contrast matters: in Kyorugi, the ecosystem includes physical devices and sensors, while in Poomsae, the judge types in subjective values. Ergonomics is not an aesthetic detail: it is an error rate, especially when there are two performers and narrow time windows to confirm entries.

3) The rule allows for subjectivity and “resolves” it with arithmetic

The heart of the WT rules (and their national adaptations) maintains the 10-point structure: 4.0 Accuracy and 6.0 Presentation, with subcriteria such as power/speed, rhythm/tempo, and expression of energy. The statistical method (average without the highest and lowest) does not correct for observation biases, distractions, or aesthetic preferences; it merely mitigates outliers.

4) Is there an alternative? Yes: computer vision + AI, today

Scientific literature and recent developments show robust models of action recognition, posture estimation, and temporal analysis that identify techniques and measure angles from multiple cameras without losing performance due to the point of view. Several studies propose to objectify Poomsae in order to address the problem of consistency and fairness in evaluation.

Even in competitive taekwondo, there is already an AI pipeline to classify actions and assist judges, with generalization to different styles and camera angles. Although the focus of some work is on sparring, the methodology (pose detection, temporality, confidence thresholds, and human review triggering) is transferable to Poomsae.

What would a serious pilot for Poomsae look like?

Setup: 4 synchronized 4K cameras (front, rear, and diagonal).

Model: 2D/3D pose + joint angle reconstruction and detection of predefined technical sequences.

Standardized metrics: angular range, stability (balance), transition times, continuity/rhythm, and power (angular velocities).

Hybrid score: the system calculates accuracy; presentation remains with judges but with visual assistance/alerts (rhythm shifts, pauses).

Confidence thresholds: if the model is <90% confidence, it requires human review; if it is ≥90%, it suggests a score with traceability (clip and markers).

5) Foreseeable objections… and responses

“AI takes the ‘art’ out of Poomsae.”
No: it separates the measurable (angles, synchrony) from the expressive (character, kiap, fluidity), which can remain in human hands. Hybridization reduces arbitrariness without “killing” aesthetics.

“Each country trains different styles.”
That is precisely why tolerances and acceptable ranges per school are used in the technical parameters; AI does not force a single “mold,” it only detects deviations from recognized patterns.

“Implementing it is expensive.”
The costs of a ring with four cameras and local computing are now affordable compared to the damage to the sport’s reputation and the cumulative expense of protests/reviews. (See AI adoptions in TKD performance analysis).

6) What WT and organizers should change (starting tomorrow)

Eliminate simultaneous face-to-face competition in phases where there is a high impact on medals; return to individual execution when the schedule allows.

Interfaces with physical feedback for judges (button panels/push buttons or “haptic tablets”), especially in mass categories.

“Single view” protocol: if simultaneity is maintained, assign pairs of judges dedicated to each athlete and a third “comparative” judge only for Presentation.

Post-competition transparency: publish a breakdown of sub-metrics (Accuracy/Presentation) and ranges per judge; today, many regulations determine the average, but do not explain how that number was arrived at.

Multi-camera AI pilot in selected G-rank events with external auditing and publication of typical errors (false positives/negatives).

7) Legal and integrity note

This analysis questions a system, it does not accuse individuals. The confidential testimonies quoted responds to the public interest of improving competitive fairness. World Taekwondo already recognizes in its frameworks that Poomsae is scored on Accuracy (4.0) + Presentation (6.0); the discussion is how to reduce operational arbitrariness and inter-judge variability, especially in simultaneous formats and with suboptimal interfaces.

Appendix: How scoring works today

•⁠ ⁠Total 10.0 = Accuracy (4.0) + Presentation (6.0).
•⁠ ⁠Presentation weighs power/speed, rhythm/tempo, and expression of energy.
•⁠ ⁠Scores are averaged, eliminating maximum and minimum; does not correct for observation biases.

About The Author

MASTKD

MASTKD: Media About Sports Taekwondo. Worldwide Leader on Taekwondo Information. «SINCE 2003»

See author's posts