Interest in Soft Body Methods

Hi all,

I am interested in verifying and understanding some of the FEM methods in PyBullet and any potential shortcomings relating to the accuracy of the actual algorithm itself. Does anyone have a short overview of the architecture/methods relating to FEM/Soft bodies in PyBullet?

Thanks!

@xhan0619

There's a brief overview of the algorithm here:

bullet3/src/BulletSoftBody/btDeformableMultiBodyDynamicsWorld.cpp

Lines 16 to 30 in 39b8de7

    
           /* ====== Overview of the Deformable Algorithm ====== */ 
        
           /* 
        
           A single step of the deformable body simulation contains the following main components: 
        
           Call internalStepSimulation multiple times, to achieve 240Hz (4 steps of 60Hz). 
        
           1. Deformable maintaintenance of rest lengths and volume preservation. Forces only depend on position: Update velocity to a temporary state v_{n+1}^* = v_n + explicit_force * dt / mass, where explicit forces include gravity and elastic forces. 
        
           2. Detect discrete collisions between rigid and deformable bodies at position x_{n+1}^* = x_n + dt * v_{n+1}^*. 
        
           3a. Solve all constraints, including LCP. Contact, position correction due to numerical drift, friction, and anchors for deformable. 
        
           3b. 5 Newton steps (multiple step). Conjugent Gradient solves linear system. Deformable Damping: Then velocities of deformable bodies v_{n+1} are solved in 
        
                   M(v_{n+1} - v_{n+1}^*) = damping_force * dt / mass, 
        
              by a conjugate gradient solver, where the damping force is implicit and depends on v_{n+1}. 
        
              Make sure contact constraints are not violated in step b by performing velocity projections as in the paper by Baraff and Witkin https://www.cs.cmu.edu/~baraff/papers/sig98.pdf. Dynamic frictions are treated as a force and added to the rhs of the CG solve, whereas static frictions are treated as constraints similar to contact. 
        
           4. Position is updated via x_{n+1} = x_n + dt * v_{n+1}.

It also contains the relevant references.

	/* ====== Overview of the Deformable Algorithm ====== */

	/*
	A single step of the deformable body simulation contains the following main components:
	Call internalStepSimulation multiple times, to achieve 240Hz (4 steps of 60Hz).
	1. Deformable maintaintenance of rest lengths and volume preservation. Forces only depend on position: Update velocity to a temporary state v_{n+1}^* = v_n + explicit_force * dt / mass, where explicit forces include gravity and elastic forces.
	2. Detect discrete collisions between rigid and deformable bodies at position x_{n+1}^* = x_n + dt * v_{n+1}^*.

	3a. Solve all constraints, including LCP. Contact, position correction due to numerical drift, friction, and anchors for deformable.

	3b. 5 Newton steps (multiple step). Conjugent Gradient solves linear system. Deformable Damping: Then velocities of deformable bodies v_{n+1} are solved in
	M(v_{n+1} - v_{n+1}^) = damping_force dt / mass,
	by a conjugate gradient solver, where the damping force is implicit and depends on v_{n+1}.
	Make sure contact constraints are not violated in step b by performing velocity projections as in the paper by Baraff and Witkin https://www.cs.cmu.edu/~baraff/papers/sig98.pdf. Dynamic frictions are treated as a force and added to the rhs of the CG solve, whereas static frictions are treated as constraints similar to contact.
	4. Position is updated via x_{n+1} = x_n + dt * v_{n+1}.