Aido Luthfi Al Hakim, Marsanda Izzah Nabilah, Wilda Imama Sabilla, Ariadi Ririd
We address the challenge of reproducing Indonesian traditional dance on a small humanoid with limited joints and frequent self-occlusions. Our monocular pose-to-robot pipeline for ROBOTIS-OP3 combines absolute 3D pose estimation (MeTRAbs), temporal refinement (DeciWatch), and a lightweight trigonometric inverse kinematics layer with coordinate normalization, scaling, and limit clamping. Trained on AIST++ with SPIN labels, the refined poses reduce mean per-joint position error (MPJPE) from 107 mm to 64.98 mm and yield smoother trajectories suited for hardware execution. The system focuses on head and arm imitation, supports culture-aware demonstrations, and can be ported to other small humanoids by updating link lengths and limits. © 2025 IEEE.
Politeknik Negeri Malang (Polinema), Department of Information Technology, Malang, Indonesia; Politeknik Negeri Malang (Polinema), Department of Electrical Engineering, Malang, Indonesia