Search papers, labs, and topics across Lattice.
This study investigates the intrinsic geometry of diffusion models through the lens of constrained inverse kinematics (IK), utilizing analytically tractable task-space constraints to evaluate the geometric structures learned by the models. By training a conditional diffusion model on two robotic arms (6-DoF UR5 and 7-DoF Franka) across various constraint families, the authors demonstrate that the intrinsic dimensions recovered align with the known degrees of freedom of the corresponding constraint manifolds. Additionally, the findings show that linear interpolation in the latent space produces solutions that adhere closely to the constraint manifolds, suggesting that the models effectively capture more than just intrinsic dimensionality, reflecting a deeper understanding of the geometric relationships involved.
The intrinsic dimensions learned by diffusion models align perfectly with the analytical degrees of freedom of constraint manifolds, revealing a surprising depth of geometric understanding in these models.
Recent studies suggest that diffusion models can recover geometric structure in the data manifolds they are trained on, yet the supporting evidence has so far come mostly from natural-image data, where the underlying geometry itself is unknown. We study this question in a setting where the geometry is analytically tractable: constrained inverse kinematics (IK). Each task-space constraint defines a configuration-space manifold with known intrinsic dimension, giving direct ground truth for evaluating the geometry learned by the model. For each of the 6-DoF UR5 and 7-DoF Franka, we train a single conditional diffusion model across seven constraint families, spanning solution manifolds from discrete IK branches to self-motion manifolds. Our empirical results reveal that the intrinsic dimension recovered from the model's score function matches the analytical degrees of freedom of the corresponding constraint manifold across both robots. Moreover, linear interpolation in the latent space leads to generated solutions that remain close to the appropriate constraint manifold, indicating that the learned representation further captures geometric structure of the constraint family beyond intrinsic dimension alone. Constrained IK therefore offers a controlled setting for studying the intrinsic geometry learned by diffusion models.
