WebGPU Implementation of Augmented Vertex Block Descent

This is an experimental WebGPU rigid-body / soft-body physics prototype. At its core is an implementation that ports an AVBD (Augmented Vertex Block Descent) style solver to WebGPU, released alongside a live demo.

Installation and running are simple.

• npm install
• npm run dev
• production build: npm run build

However, this is still an early proof of concept: it's not a plug-and-play module yet, and browser support is currently Chrome only.

The implementation follows the flow of Algorithm 1 from the AVBD paper fairly directly.

• collision detection: performs collision detection at the current position x^t at the start of the time step.
• broad phase: builds and traverses an LBVH to find candidate body pairs.
• narrow phase: builds manifolds and maintains per-contact state for warm-start and friction handling.
• per-body constraint list: gathers the constraints each body should reference based on contacts, joints, and springs.
• coloring: uses greedy coloring to prepare bodies of the same color for parallel processing.
• primal init: builds the inertial target y and initializes primal state.
• colored primal solve: during iterative passes, solves bodies color by color and applies the AVBD primal update.
• dual/stiffness update: updates dual variables and stiffness after the primal sweep.
• finalize velocities: reconstructs velocity from the final state.

The documentation emphasizes that while the current pipeline closely resembles the paper, the double-buffered position updates needed for rare same-color conflicts have not yet been implemented as-is, and the current path is essentially an in-place colored body solve.

In other words, this is less a finished library and more an experimental implementation validating AVBD on WebGPU, making it a meaningful technical demo for developers interested in physics simulation and web graphics.