We propose a split-latent hierarchical diffusion framework that decomposes a 3D scene into a fixed latent — encoding deterministic geometric and semantic priors — and a diffused latent — capturing stochastic, strategy-dependent context via a Global Conditional Diffusion guided by RGB. A Local Conditional Diffusion then predicts dense, point-wise affordance and grasp fields, regularized by a Harmonic Grasp Field for smooth, obstacle-aware grasp manifolds. The result is grasp generation that is geometrically precise yet contextually diverse in cluttered, partially observed scenes.
Method
Split-Latent Diffusion
A hierarchical architecture disentangling deterministic geometric priors from stochastic strategic uncertainty for robust reasoning in clutter.
Global + Local Diffusion
A conditioning mechanism fusing global stochastic context with local point-wise geometry to produce uncertainty-aware grasp fields.
Harmonic Grasp Field
A principled coupling of diffusion and differential geometry, yielding smooth, continuous, obstacle-aware dexterous grasp manifolds.
Results
93.1
Dexterous GSR
(DexGraspNet-2.0, Dense)
79.0
Gripper GSR
(GraspClutter6D, Pile-15)
79.8
AP Seen
(GraspNet-1Billion)
98.2
Pick Success
(A² sim, Unseen)
State-of-the-art across dexterous grasping, gripper grasping, and policy-rollout benchmarks. See the paper for full tables.
Get the Code
The full implementation — encoder pretraining, hierarchical diffusion, Harmonic Grasp Field, and grasp VAE — is on GitHub and runs end-to-end: