Osama Al-Shammari

Artificial Intelligence Specialization · Cohort 2026

Physics-Informed Continuous Autoencoding for Compact 3D Human Motion

Digital human motion contains many body positions for every second, making it expensive for AI models to store and process. Normal compression can lose important details, causing the reconstructed character to drift, slide its feet, or move incorrectly. PIAE compresses every eight frames while preserving the character’s path, turning, height, and foot contact. When tested on 4,210 HumanML3D motions, PIAE reconstructed joint positions with an average error of only 8.94 mm—less than one centimetre—while producing almost zero final-position drift. Under the same evaluation, MotionGPT and MLD produced errors of 172.09 mm and 104.41 mm, respectively. PIAE also achieved more accurate movement speed and substantially less foot sliding.

Read the final report (PDF)