Diffusion models are believed to recover the geometric structure of their training data, but that claim has mostly been tested on natural images, where the true geometry is unknown. The authors of this paper instead use constrained inverse kinematics, where each task-space constraint defines a configuration-space manifold with a known intrinsic dimension, providing explicit ground truth.
They trained a single conditional diffusion model per robot—a 6-DoF UR5 and a 7-DoF Franka—across seven constraint families, including discrete IK branches and self-motion manifolds. The intrinsic dimension estimated from the model's score function matched the analytical degrees of freedom for every family on both robots.
The model also appeared to preserve structure beyond dimension: linear interpolation in latent space produced solutions that remained close to the correct constraint manifold. The authors argue that constrained IK is a controlled, analytically tractable setting for studying the intrinsic geometry learned by diffusion models, complementing less interpretable image-domain evidence.