SofaDiff

Plugin for differentiable physics in SOFA.

Installation

Cf installation of SOFA plugins.

Note: the examples may rely on a change in SpringForceField that has not been pushed on the master branch of SOFA yet (TODO).

Features

Supports:

• Differentiation of static and quasi-static simulations
• Propagation of the gradient back to parameters in the force fields
  • Requires "augmenting" the force field, cf the ParameterizedSpringForceField example
  • Or implementation from scratch in Python (cf below)
• Gradient-based optimization of said parameters
• Implementation of components in Python:
  • Optimization algorithm (e.g. gradient descent based on JAX)
  • Parameterized force fields (e.g. force field leveraging JAX autodiff)

Will support:

• Differentiation of dynamic simulations
• Propagation of the gradient back to parameters in projective constraints
• Differentiation through Lagrange constraints & contacts