Experimental WebGPU Port of “Augmented Vertex Block Descent” Lands on GitHub

What dropped

A new repository called webphysics is out, offering a WebGPU-based prototype of rigid- and soft-body simulation that, it has been reported, centers on an AVBD-style (Augmented Vertex Block Descent) solver. The project includes a teaser video and a live demo, plus simple dev commands (npm install; npm run dev) so you can spin it up locally. It’s an early proof of concept — browser support is limited (Chrome only, for now) and the author warns it’s not plug-and-play. So don’t expect a polished engine just yet. Excited? You should be. This is the sort of tinkering that pushes the web forward.

How it works under the hood

The repo closely follows Algorithm 1 from the AVBD paper, mapping paper sections to concrete GPU code: collision detection flows into LBVH-based broad phase (src/physics/gpu/broadPhase.ts, src/lvbh/GPULBVHBuilder.ts), then narrow-phase contact and warm-start persistence (src/physics/gpu/contactGeneration.ts, src/physics/gpu/contactRecord.ts). Per-body constraint gathers, greedy coloring, inertial target initialization, colored primal solves, dual/stiffness updates, and final velocity reconstruction are all implemented across src/physics/gpu/avbdState.ts and src/physics/PhysicsEngine.ts. Short version: collision detection → coloring → inertial/primal init → repeated colored primal solves → dual updates → finalize velocities. The mapping between paper and code is impressively explicit; you can trace algorithmic steps to filenames and functions.

Why it matters (and what’s next)

Bringing a modern augmented-Lagrangian solver like AVBD into WebGPU is a small technical revolution. Why run heavy physics on the server or in native-only engines when browsers are finally gaining the GPU APIs to do real work? That said, this is experimental: stability, performance, and ergonomics are on the to-do list. One important implementation difference — the paper’s double-buffered position update is reportedly handled differently here — and that divergence will be a key story to watch as the project matures. If you care about high-fidelity, real-time physics on the web (games, simulators, interactive art), this repo is worth a bookmark and maybe a star. Want to poke under the hood? The code’s laid out for exactly that.

Sources: github.com/jure, Hacker News