Pencak Silat sebagai seni bela diri tradisional Indonesia memerlukan ketelitian dalam pembentukan gerakan dasar yang benar. Namun, evaluasi gerakan masih bergantung pada penilaian subjektif pelatih dan terbatas pada latihan tatap muka. Penelitian ini bertujuan merancang dan membangun aplikasi Android SilatPRO untuk mengevaluasi gerakan dasar Pencak Silat kategori pukulan dan tangkisan secara objektif dan real-time. Metode yang digunakan meliputi deteksi pose tubuh menggunakan YOLOv8n-Pose berbasis Convolutional Neural Network (CNN) untuk mengekstraksi 17 keypoint, kemudian diklasifikasikan menggunakan Multilayer Perceptron (MLP) dan dievaluasi dengan pendekatan rule-based. Dataset diperoleh dari video gerakan yang direkam langsung di Perguruan BIMA SUCI, lalu dibagi menjadi 70% data latih, 15% validasi, dan 15% pengujian. Gerakan yang dievaluasi terdiri dari dua kategori dengan variasi kanan dan kiri sehingga menghasilkan 10 kelas gerakan. Hasil pengujian menunjukkan model MLP mencapai akurasi 100% dengan nilai precision, recall, dan F1-score sebesar 1,00. Black Box Testing menunjukkan seluruh fitur berjalan sesuai kebutuhan, sedangkan User Acceptance Testing (UAT) melibatkan 25 responden dan menghasilkan kelayakan 87,92%. Dengan demikian, SilatPRO mampu mengevaluasi gerakan secara real-time melalui klasifikasi gerakan, skor penilaian, dan umpan balik otomatis sebagai media latihan mandiri. Aplikasi ini juga dapat membantu pesilat mengetahui kesalahan gerakan secara lebih cepat, objektif, dan konsisten tanpa selalu didampingi pelatih.

Pencak Silat, as a traditional Indonesian martial art, requires precision in forming correct basic movements. However, movement evaluation still depends on trainers’ subjective judgment and is limited to face-to-face practice sessions. This study aims to design and develop an Android application named SilatPRO to evaluate basic Pencak Silat movements in punch and block categories objectively and in real-time. The method includes body pose detection using YOLOv8n-Pose based on Convolutional Neural Network (CNN) to extract 17 body keypoints, which are then classified using Multilayer Perceptron (MLP) and evaluated through a rule-based approach. The dataset was obtained from movement videos recorded directly at Perguruan BIMA SUCI and divided into 70% training data, 15% validation data, and 15% testing data. The evaluated movements consist of two categories with right and left variations, resulting in 10 movement classes. The testing results show that the MLP model achieved 100% accuracy with precision, recall, and F1-score values of 1.00. Black Box Testing showed that all features worked according to system requirements, while User Acceptance Testing (UAT), involving 25 respondents, produced a feasibility score of 87.92%. Therefore, SilatPRO can evaluate movements in real-time through movement classification, scoring, and automatic feedback as an independent training medium. This application can also help learners identify movement errors more quickly, objectively, and consistently without always being accompanied by a trainer.