Modern sensing can track the motion of physical systems, but the forces and mechanical response that drive that motion often remain unobserved. Recovering these hidden quantities from discretely sampled trajectories is a long-standing challenge, and it becomes especially difficult when the sampling is coarse — that is, when the time steps between observations are large.
A new paper on arXiv tackles this problem with graph networks, a class of models well suited to relational data. The authors propose learning coarse-step dynamics directly from trajectory data while simultaneously inferring the internal mechanical response of the system. Their approach aims to bridge the gap between what is measured and what is physically happening beneath the surface.
The work is still at the preprint stage, and the abstract provided does not include experimental results or comparisons to existing methods. However, the framing suggests a practical route for extracting force information from sparse sensor data, which could benefit fields ranging from robotics to structural monitoring. As with any preprint, the findings await peer review and further validation.