▶ Live demo: https://astaroverov.github.io/exp_webgpu_tf/ (needs a WebGPU browser — Chrome/Edge desktop. Click once to enable sound.)
A from-scratch browser game where you drive a tank against an opponent that was trained with reinforcement learning — and everything (physics, GPU rendering, and the neural-net training itself) runs client-side in a single tab, with no backend. A personal hobby project to see how far the browser can actually be pushed.
Three things built by hand, no engines or game frameworks:
- My own ECS game engine — ~50 components / ~40 systems on
bitecs, with a 2D Rapier physics backend. Strictly data-oriented: behavior is a query over components, never a branch inside a system. Runs headless and deterministic so it can be stepped as fast as possible for training. - My own WebGPU renderer — hand-written WGSL + SDF shape passes and a screen-space Radiance Cascades global-illumination lighting pass. No Three.js/pixi.
- From-scratch PPO (Proximal Policy Optimization) on TensorFlow.js that actually trains the agent you fight.
A top-down desert combat prototype. Vehicles are composed from parts bolted onto a compound Rapier body (hull, turret, wheels, tracks), so they take localized damage and shed debris instead of being one rigid blob. Multiple weapon families share a single damage pipeline: ballistic guns, rockets, an EMP gun, and continuous flame / frost streams with damage-over-time and slow effects — plus shields, repair, scoring, and destructible terrain.
A distributed actor/learner RL pipeline running entirely across Web Workers: actors run the headless sim + inference on the WASM TF backend, learners train on the WebGPU backend, and the main tab shows a live visualizer + metrics dashboard. It uses action masking, an opponent population (self-play + frozen past selves + scripted baselines), and a curriculum. The same headless engine serves both the trainer and the playable game — the opponent in the demo is one of these trained networks, run through the exact training inference path.
TypeScript · Web Workers · WebGPU + WGSL (custom renderer) · bitecs (ECS) ·
@dimforge/rapier2d (physics) · TensorFlow.js (webgpu + wasm) · Web Audio.
Active personal experiment — APIs, package names, and the game itself change freely as I explore. Shared as a record of what I'm tinkering with, not as a product.