DiffPD: Differentiable Projective Dynamics

We present a novel, fast differentiable simulator for soft-body learning and control applications. Existing simulators can be classified into two categories based on their time integration methods: Simulators using explicit timestepping schemes require tiny timesteps to avoid numerical instabilities in gradient computation, implicit typically compute gradients by employing the adjoint method solving expensive linearized dynamics. Inspired Projective Dynamics ( PD ), we Differentiable DiffPD an efficient with integration. The key idea is speed up backpropagation exploiting prefactorized Cholesky decomposition forward simulation. In terms of contact handling, supports types contacts: penalty-based model describing friction forces complementarity-based enforcing non-penetration conditions static friction. evaluate performance observe it 4–19 times faster compared standard Newton’s various applications including system identification, inverse design problems, trajectory optimization, closed-loop control. also apply reality-to-simulation real-to-sim ) example collisions show its capability reconstructing digital twin real-world scenes.