Ramadhan, Muhammad Raihan (2026) Neural Network-Based Ball Trajectory Control for Autonomous Soccer Robot Goalkicking. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6022241102-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6022241102-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (1MB) |
Abstract
Autonomous soccer robots competing in KRSBI Division B face a critical challenge: achieving precise ball trajectory control within the 15-minute field access window provided before competitions. Traditional manual tuning is operationally infeasible under this constraint and fails to capture non-linear solenoid actuator dynamics. This thesis presents a rapid adaptation framework combining sparse sampling, synthetic data generation, and neural network modeling. The sparse sampling protocol reduces calibration from 23 to 9 points, enabling completion within the time constraint. Random Forest-based generation reconstructs the full workspace with R2 = 0.96 (MAE = 8.39 counts), while Leave-One-Out Cross-Validation confirms generalization with R2 = 0.7695 (MAE = 16.69 counts) from sparse real data. The Multi-Layer Perceptron (2→10→10→1) achieves R2 = 0.9966 (MAE = 2.35 counts), representing 3× improvement over Random Forest and 8× over Polynomial Regression. The model produces a smooth control surface essential for holonomic motion integration, unlike Random Forest’s blocky discontinuities or Polynomial Regression’s systematic bias. Ablation studies demonstrate exceptional robustness with Error Amplifi cation Ratio of 0.09 (10× noise dampening). At 5 cm localization uncertainty, prediction error remains below 0.07 mm with >99% within tolerance. Look Up Table optimization achieves 17.10 µs inference (56% reduction). Field validation confirms 100% valid goal rate despite systematic localization bias, validating deployment readiness for competitive environments.
====================================================================================================================
Robot sepak bola otonom yang berkompetisi di KRSBI Divisi B menghadapi tantangan kritis: mencapai kontrol lintasan bola yang presisi dalam jendela akses lapangan 15 menit yang diberikan sebelum kompetisi. Metode tuning manual tradisional tidak layak secara operasional dengan kendala waktu ini dan gagal menangkap dinamika aktuator solenoid yang non-linear. Penelitian ini menghadirkan kerangka kerja adaptasi cepat yang meng gabungkan sparse sampling, pembangkitan data sintetis, dan pemodelan neural network. Protokol sparse sampling mengurangi kalibrasi dari 23 menjadi 9 titik, memungkinkan penyelesaian dalam kendala waktu tersebut. Pembangkitan berbasis Random Forest merekonstruksi workspace lengkap dengan R2 = 0.96 (MAE = 8.39 counts), sedangkan Leave-One-Out Cross Validation mengonfirmasi generalisasi dengan R2 = 0.7695 (MAE = 16.69 counts) dari data riil yang sparse. Multi-Layer Perceptron (2→10→10→1) mencapai R2 = 0.9966 (MAE = 2.35 counts), yang merepresentasikan peningkatan 3× dibanding Random Forest dan 8× dibanding Polynomial Regression. Model menghasilkan control surface yang halus dan esensial untuk integrasi motion holonomik, tidak seperti diskontinuitas Random Forest yang blok-blok atau bias sistematis Polynomial Regression. Studi ablasi mendemonstrasikan robustness yang luar biasa dengan Error Amplification Ratio sebesar 0.09 (peredaman noise 10×). Pada ketidakpastian lokalisasi 5 cm, error prediksi tetap di bawah 0.07 mm dengan >99% dalam toleransi. Optimasi Look-Up Table mencapai inferensi 17.10 µs (reduksi 56%). Validasi lapangan mengonfirmasi 100% valid goal rate meskipun terdapat bias lokalisasi sistematis, memvalidasi kesiapan deployment untuk lingkungan kompetitif.
Actions (login required)
 |
View Item |