- Public site live at ignitionai.dev
- Single Next.js deployment serving the landing, the docs, and the demos
- Brand: animated flame GIF logo, indigo palette, 3 custom inline SVG feature diagrams (ONNX pipeline, algorithm cards, R3F integration)
- Flame favicon across every page (
app/icon.svg) and every demo
- 20-page Nextra documentation site under
/docs - Introduction + Quickstart (7-line CartPole example, real package imports)
- Verbose algorithm pages for DQN, PPO, and Q-Table with source-cited defaults and failure-mode recipes
- One "How it works" page per backend package (
core,backend-tfjs,backend-onnx,storage) with annotated source walkthroughs - React Three Fiber page explaining the training-loop / render-loop split
- 7 step-by-step tutorials: GridWorld, CartPole observations, MountainCar reward shaping, CartPole 3D, Car Circuit, ONNX → Unity, Drone Navigation
- 5 existing demos embedded as static routes under
/demos/<slug>/: GridWorld, CartPole, MountainCar, CartPole 3D, Car Circuit ← IgnitionAIback link in every demo header- Prebuild pipeline that builds the library packages and each Vite demo with the right base path before the Next.js build
@ignitionai/demo-drone-navigation— a quadcopter that learns to fly to moving target points- Hand-rolled rigid-body physics: gravity, drag, torque from asymmetric thrust, semi-implicit Euler integration at 50 Hz
- 8-action discretization (hover / forward / back / left / right / yaw) so DQN can handle 6-DOF flight
- Reward shaping: distance + progress delta + anti-spin + capture bonus + crash penalty
- Tuned defaults:
hiddenLayers: [64, 64],minEpsilon: 0.05,epsilonDecay: 0.998— visible hovers around episode 200–400
- Replaced the
+1 per on-track stepreward with dense progress shaping (progressDelta × 300+ alignment + centerline). The old reward treated "spinning in place" and "making laps" as equally valuable.
TrainingEnvinterface — implement 5 methods + actions to define your environmentInferenceEnvinterface — minimal observe + step for production deploymentIgnitionEnv— training loop withtrain(),infer(),stop(),reset()- Auto-config: deduces inputSize/actionSize from your environment
- Algorithm switching:
env.train('dqn'),env.train('ppo'),env.train('qtable') - Greedy mode for inference:
getAction(obs, greedy=true)
- DQN — Deep Q-Network with replay buffer, target network, epsilon-greedy
- PPO — Proximal Policy Optimization with Actor-Critic, GAE-lambda, entropy bonus
- Q-Table — Tabular Q-Learning with mixed-radix state discretization
IgnitionEnvTFJS— IgnitionEnv with TF.js agent factories baked in- Sensible defaults per algorithm — zero config required
- Configurable TF.js backend selector (WebGPU > WebGL > WASM > CPU)
OnnxAgent— inference-only agent for ONNX Runtime- TF.js → ONNX exporter with Python conversion script
- HuggingFace Hub loader for .onnx models
ModelStorageProviderinterface (save/load/list/delete/exists)- HuggingFace Hub provider with retry and Zod config validation
GridWorldEnv— NxN grid pathfinding (discrete)CartPoleEnv— Pole balancing with Euler physics (continuous state)MountainCarEnv— Momentum-based hill climbing (sparse reward)- All implement
TrainingEnv— plug and play withIgnitionEnv