Train RL agents in your browser
IgnitionAI is the ML-Agents of the JavaScript creative ecosystem. Describe your game world, call env.train(), and deploy anywhere via ONNX. No Python, no server, no GPU cluster.
Train your first agent in 7 lines
No neural network code. No hyperparameter tuning. No config files. Describe your world, call train(), and the framework does the rest.
import { IgnitionEnvTFJS } from '@ignitionai/backend-tfjs';
import { CartPoleEnv } from '@ignitionai/environments';
const cartpole = new CartPoleEnv();
const env = new IgnitionEnvTFJS(cartpole);
env.train('dqn'); // Zero config. It just works.
// env.infer(); // Switch to inference after training.
// env.setSpeed(50); // Turbo training (50x faster).That's it. The agent learns. The pole stays up.
Reinforcement learning that fits your stack
Describe your world in a class, call env.train(), and watch your agent learn in real time. No Python, no server, no GPU cluster.
Everything you need to ship RL.
Three algorithms, auto-configuration, ONNX export, HuggingFace storage, and five working demos. All TypeScript, all open source.
Train → ONNX → deploy anywhere
Train in the browser with TensorFlow.js, then export to ONNX and deploy in Unity (Sentis), Unreal (NNE), Python, C++, or edge devices. One pipeline from prototype to production.
Three algorithms, one API
DQN, PPO, and tabular Q-Learning. Switch with one word: env.train('ppo'). Override hyperparameters only when you need fine control.
R3F-first
Pair it with your Three.js scene — training loop runs independently of the render loop.
Auto-configuration
Input size deduced from observe(), action size from actions.length. You never touch network code.
WebGPU acceleration
TensorFlow.js auto-selects WebGPU → WebGL → WASM → CPU. No CUDA, no install.
Turbo training
env.setSpeed(50) — 50× faster convergence. Drop back to 1× to watch the agent play.
HuggingFace storage
Save and load trained models from HF Hub with one line. Share weights like a dataset.
Strict TypeScript
No any, Zod-validated configs, 184+ tests, CI/CD. Production-ready from day one.
Five working demos
GridWorld, CartPole, MountainCar, CartPole 3D, and Car Circuit. Clone, run, learn.
Train an agent in your browser
Pick an algorithm, hit Train, and watch the reward climb in real time. This is a simulation — open a live demo below to run real RL.
Watch agents learn in real time
Six interactive environments showing IgnitionAI in action. Each runs in your browser, trains locally, and demonstrates a different RL challenge.
Maze
A low-poly humanoid trains in a Rapier 3D arena with prefab walls, sensor rays, doors, and moving hazards.
CartPole
Classic pole-balancing benchmark with Euler physics. Converges fast.
MountainCar
Sparse reward challenge. Agent discovers the counterintuitive momentum strategy.
CartPole 3D
Classic CartPole rendered in 3D with React Three Fiber. Metallic materials and shadows.
Car Circuit
A 3D car learns to drive an oval circuit. Chase cam, minimap, trail, 1×–50× speed.
Drone Navigation
A quadcopter learns to fly to moving target points. Rigid-body physics, 8 thrust combos, real gravity and torque. The hero demo.
Maze 3D
A robot navigates a 3D maze to find keys, avoid traps, and reach the exit. Raycast observations, pathfinding challenge.
What's new
IgnitionAI ships fast. Here's the latest release — everything else lives on the full changelog.
- Landing & sitePublic site live at [ignitionai.dev](https://ignitionai.dev)
- Landing & siteSingle Next.js deployment serving the landing, the docs, and the demos
- Landing & siteBrand: animated flame GIF logo, indigo palette, 3 custom inline SVG feature diagrams (ONNX pipeline, algorithm cards, R3F integration)
- Landing & siteFlame favicon across every page (`app/icon.svg`) and every demo
- Documentation20-page Nextra documentation site under `/docs`
Ready to train your first agent?
One install command, one TrainingEnv class, one call to env.train(). Your agent learns in your browser, then deploys anywhere via ONNX.