Abstract
Tracking and forecasting the rotation of objects is fundamental in computervision and robotics, yet SO(3) extrapolation remains challenging as (1) sensorobservations can be noisy and sparse, (2) motion patterns can be governed bycomplex dynamics, and (3) application settings can demand long-termforecasting. This work proposes modeling continuous-time rotational objectdynamics on $SO(3)$ using Neural Controlled Differential Equations guided bySavitzky-Golay paths. Unlike existing methods that rely on simplified motionassumptions, our method learns a general latent dynamical system of theunderlying object trajectory while respecting the geometric structure ofrotations. Experimental results on real-world data demonstrate compellingforecasting capabilities compared to existing approaches.