IgnitionAI

IgnitionAI is a reinforcement learning framework for JavaScript developers. Describe your game world as a class, call env.train('dqn'), and watch a neural network learn to play it — live, in your browser, with no Python, no server, and no GPU cluster.

The framework is the JavaScript answer to Unity’s ML-Agents. It’s built for the people who already ship creative work in the JS/TS stack: Three.js authors, React Three Fiber devs, generative-art hackers, game-jam builders, and anyone who wants to add a learning agent to an interactive scene without leaving their existing toolchain.

Why IgnitionAI exists

Unity has ML-Agents . Python has Stable Baselines, RLlib, and CleanRL. JavaScript had nothing comparable — until now.

If you work in the JS creative ecosystem, your options for adding RL to your projects have historically been:

Write everything in Python and bridge it back via WebSockets. Works, but breaks your toolchain and adds a server to a project that was supposed to run entirely in the browser.
Hand-roll a training loop on top of TensorFlow.js. Possible, but you’re now maintaining your own RL library on top of your game.
Give up and use a scripted AI. Most people pick this one.

IgnitionAI closes that gap. You implement a handful of methods describing your world, you pick an algorithm, and the framework handles the neural network, the replay buffer, the exploration schedule, the training loop, and the backend selection.

What makes it different

Zero config. Implement five methods (observe, step, reward, done, reset) plus an actions list. The framework deduces the network shape from your first observation. You never touch neural-network code unless you want to.
Browser-native. TensorFlow.js auto-selects WebGPU → WebGL → WASM → CPU. No install, no CUDA, no server. Everything runs where your game runs.
Train → Deploy pipeline. Train in the browser, export to ONNX, and deploy the .onnx file in Unity (Sentis), Unreal (NNE), Python, C++, or edge devices. IgnitionAI is a serious tool, not a toy.
R3F-first. The training loop is decoupled from the render loop, so your <Canvas> stays at 60 fps while the agent learns in the background. You can also call env.setSpeed(50) to fast-forward training and then drop back to real-time to watch the policy play.
Production-ready. TypeScript strict mode, Zod validation, 184+ tests, CI/CD, modular monorepo. The framework is small but not fragile.

How it compares

	IgnitionAI	Unity ML-Agents	Python (SB3 / RLlib)
Language	TypeScript / JavaScript	C# + Python	Python
Where training runs	Browser	Unity editor + Python sidecar	Python process
Install	`npm install`	Unity + Python toolchain	pip + CUDA
Primary audience	JS/TS creative devs	Unity game devs	ML researchers
Deploy target	Browser, ONNX anywhere	Sentis / Barracuda	Production servers
GPU required	No	Optional	Usually yes

None of these is strictly better than the others — they target different audiences. If you already work in JS, IgnitionAI meets you where you are.

Install


npm install @ignitionai/core @ignitionai/backend-tfjs @ignitionai/environments

Three packages to get started. core contains the training loop and the TrainingEnv interface. backend-tfjs provides the TensorFlow.js agents (DQN, PPO, Q-Table). environments ships ready-made envs (CartPoleEnv, GridWorldEnv, MountainCarEnv) so you can train something in 30 seconds. Drop environments from the list if you’d rather write your own env from scratch — the tutorials walk through that exact flow.

Where to go next

Quickstart — train your first agent in under 10 lines. Start here.
Algorithms — the “verbose” pedagogical explanations of DQN, PPO, and Q-Table, with every default hyperparameter and when to use each.
How it works — the architecture: one page per backend package, with annotated walkthroughs of the source.
React Three Fiber — why IgnitionAI is built for R3F, and a complete <Canvas> + TrainingEnv example.
Tutorials — start with the GridWorld tutorial for a full step-by-step walkthrough.

Next: Quickstart →