Toward Fair Evaluation: A Unified Transformer for Objective Gymnastic Performance Analysis

In competitive gymnastics, performance evaluation and training have traditionally relied on subjective human observation, which struggles to consistently process the high-speed, complex biomechanical nuances of elite execution. While the advent of advanced motion capture and wearable sensors offers a wealth of quantitative kinematic data, there remains a critical gap in translating these raw metrics into objective, fair, and actionable coaching insights. To address this challenge, the proposed UniGym framework introduces an integrated architecture for gymnastic movement analysis and automated training recommendation, bridging the gap between raw sensor signals and qualitative coaching expertise. The system processes multi-modal kinematic data-incorporating synchronized vision-based skeletal tracking and inertial measurement units-through a unified embedding space that harmonizes high-dimensional biomechanical features with technical proficiency metrics. A core innovation of this approach is the systematic elimination of contextual biases; by employing adversarial alignment during feature encoding, the model focuses exclusively on movement quality while discarding confounding demographic variables and individual stylistic idiosyncrasies. The technical pipeline leverages weakly supervised learning to extract refined motion representations from raw, uncurated kinematic inputs, reducing the dependency on exhaustive frame-level annotations. Our transformer mechanism models each training session as a complex temporal sequence of movement events, utilizing multi-head self-attention to capture hierarchical technical patterns and long-range dependencies across diverse skill levels. To facilitate comparative analysis, the system constructs topology-aware similarity graphs from the unified embeddings, enabling large-scale motion pattern clustering that ensures consistent grouping through objective, distance-based similarity measures. For training optimization, the UniGym framework implements a generative recommendation engine that synthesizes personalized corrective strategies based on latent technical gaps identified within the movement sequences. By comparing an athlete’s performance manifold against a distilled ``expert'' representation, the model identifies non-obvious relationships across athlete groups through macro-scale interaction modeling. Evaluation on a comprehensive dataset of over ten thousand gymnastics training sequences demonstrates the model’s robustness and cross-apparatus generalizability, achieving up to a 31% improvement in the Unified Balanced Error Rate (UBER) compared to the closest state-of-the-art temporal modeling baselines, along with a 99.3% classification accuracy on expert executions. In conclusion, this study definitively establishes that the UniGym framework is a deployable, bias-free standard. It proves that integrating classification, clustering, and recommendation via adversarial debiasing strictly outperforms traditional isolated models, offering a concrete solution to the fragmentation problem in sports analytics.