PhysSkin++: Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning

PhysSkin++. Real-Time and Generalizable Physics-Based Animation via Self-Supervised Neural Skinning. By Yuanhang Lei and eleven others, from Zhejiang University, Ant Group, and three other institutions. Published on the nineteenth of October. Its code repository is public and real: nineteen source files.

Real-time physics-based animation is hard to generalize. It has to work across diverse 3D shapes and across different spatial discretizations, and doing that remains a challenge. The goal is a single learned system that can take an unfamiliar shape, in whatever form it arrives, and animate it physically in real time.

PhysSkin++ is a physics-informed framework that learns continuous neural skinning fields from static geometries, and uses them to build compact deformation subspaces. At its center is a Neural Skinning Fields Autoencoder, which predicts a shape-dependent deformation subspace without training a separate network for each object. A new shape goes in, and a compact space of ways it can deform comes out.

These subspaces are learned without any simulation trajectories to imitate. In their place, a Physics-Informed Self-Supervised Learning strategy combines skinning-field normalization with conflict-aware gradient correction, which stabilizes joint optimization under physical constraints and orthogonality constraints.

The learned representation supports efficient reduced dynamics with user control and contact interactions, and binds directly to meshes, point clouds and 3D Gaussians. For prescribed rig animation, a self-supervised complementarity objective learns complementary elastic responses, without enforcing complementarity constraints at runtime.

Experiments demonstrate that the learned skinning fields generalize across shapes, that reduced simulation runs in real time, and that the framework applies to animation that is controllable, collision-aware, and complementary to a rig. Each of those applications rests on the same learned skinning fields.

The reach comes from what is left out. No per-object network and no simulation trajectories are needed, so one learned field can turn a static shape into something that moves.