Overview
Browser-based reinforcement learning framework porting Unity ML-Agents architecture to Three.js. Enables RL agent training and deployment in WebGL environments.
Architecture
Agent-Environment Interface
Replicates Unity ML-Agents API in JavaScript:
- Observation collection
- Action application
- Reward processing
- Episode management
Training Pipeline
- Policy networks in TensorFlow.js
- Browser-based training option
- Server-side training with model export
- Policy deployment to browser
Three.js Integration
ML-Agents concepts adapted to Three.js scene graph:
- Agent entities as Three.js objects
- Physics-based environments
- Sensor raycasting
- Action space mapping
Use Cases
Browser-Based Robotics Simulation
Train manipulation and navigation policies directly in browser. No installation required for distributed training or demonstration.
Game AI Development
Rapid prototyping of game AI behaviors with visual feedback. Immediate deployment to web-based games.
Educational Applications
Accessible RL training platform requiring only web browser. Visual feedback aids understanding of policy learning dynamics.
Technical Features
- TensorFlow.js backend
- WebGL rendering via Three.js
- PPO, SAC algorithm support
- Model serialization for deployment
- Multi-agent environments
Performance
CPU training viable for simple environments. GPU acceleration via WebGL enables training of visual policies in browser.
Deployment
Trained policies bundle with Three.js application for zero-dependency web deployment. Models run entirely client-side.