Overview

This section provides an overview of GenesisLab’s design, philosophy, and architecture.

• What is GenesisLab?
  • Motivation
  • What Can You Do with GenesisLab?
  • How is GenesisLab Different?
  • Architecture at a Glance
  • Next Steps
• Architecture
  • Overview
  • Core Components
  • Term-Based Composition
  • Data Flow
  • Configuration System
  • Execution Flow
  • Extension Points
  • Next Steps
• Core Concepts
  • LabScene
  • Managers
  • Terms
  • ConfigClass
  • Scene Builder/Controller/Querier
  • Sensors
  • Terrains
  • Vectorization
  • MDP (Markov Decision Process)
  • Next Steps
• Design Philosophy
  • 1. Modularity Through Managers
  • 2. Composition Over Inheritance
  • 3. Configuration-Driven Design
  • 4. Performance First
  • 5. Sensible Defaults
  • 6. Extensibility
  • 7. Compatibility
  • 8. Clarity Over Cleverness
  • 9. Fail Fast and Loud
  • 10. Documentation as Code
  • Design Trade-offs
  • Inspirations
  • Anti-Patterns to Avoid
  • Contributing
  • Next Steps
• Frequently Asked Questions
  • General Questions
  • Installation and Setup
  • Configuration
  • Observations and Actions
  • Rewards
  • Sensors
  • Training
  • Terrains
  • Debugging
  • Sim-to-Real
  • Performance
  • Next Steps

Quick Summary

GenesisLab is a unified framework for robot learning built on Genesis physics simulation. It provides:

• Manager-based architecture for modular, extensible task definitions

• Multi-robot support with pre-configured robot platforms

• Gymnasium API integration for seamless RL library compatibility

• Advanced sensor systems with fake and native Genesis sensors

• Terrain generation with procedural generation and curriculum support

• Domain randomization through the event system

Key Concepts

Understanding these core concepts will help you work effectively with GenesisLab:

1. LabScene: The central simulation manager that coordinates all other managers

2. Managers: Specialized components handling observations, actions, rewards, etc.

3. Terms: Individual computation units within managers (e.g., observation terms, reward terms)

4. ConfigClass: Configuration system using @configclass decorators

5. Scene Builder/Controller/Querier: Patterns for interacting with the Genesis scene

Philosophy

GenesisLab is designed with several key principles:

• Modularity: Clean separation between different aspects of the learning problem

• Extensibility: Easy to add custom observations, rewards, and behaviors

• Performance: Leveraging Genesis’s speed for fast parallel training

• Simplicity: Configuration-driven design with sensible defaults

• Compatibility: Works with standard RL libraries and follows Gymnasium conventions