Skip to main content

The Simulation Stack: MuJoCo, Isaac Gym, and Isaac Lab

Duration: 50 min · Level: Intermediate · Module: 7. Simulation & Digital Twins · Focus: Isaac, MuJoCo, simulation, RL

Learning objectives

By the end of this lesson you will be able to explain and apply:

  • MuJoCo (Multi-Joint dynamics with Contact)
  • MuJoCo 3.0 (2023)
  • Isaac Gym (2021, now deprecated)
  • Isaac Lab (2024)
  • Isaac Sim 4.0

Why this matters​

Physics simulation is the training environment for G1's neural policies.

Overview​

Physics simulation is the training environment for G1's neural policies. The choice of simulator affects training speed, sim-to-real fidelity, and toolchain compatibility. The field has converged on two primary platforms: MuJoCo (DeepMind) for precision and Isaac Lab (NVIDIA) for scale.

Key concepts​

Key idea

MuJoCo (Multi-Joint dynamics with Contact): acquired by DeepMind in 2021, made free; gold standard for contact physics accuracy; used in OpenAI Gym, all major RL benchmarks

  • MuJoCo 3.0 (2023): 10-100× faster than previous versions with parallel simulation; runs 10,000+ environments on a single workstation GPU
  • Isaac Gym (2021, now deprecated): NVIDIA's GPU-parallel RL training environment; 4096+ parallel environments on a single A100; enabled ETH Zurich ANYmal locomotion results
  • Isaac Lab (2024): replaces Isaac Gym; built on Isaac Sim 4.0/Omniverse; better asset management, more realistic contact model, ray-traced rendering for visual RL
  • Isaac Sim 4.0: full USD (Universal Scene Description) workflow; Omniverse-based; photorealistic rendering for domain randomization of visual appearance
  • G1 recommended stack: Isaac Lab for locomotion training (massively parallel RL) + MuJoCo for manipulation fine-tuning (contact precision) + ROS 2 bridge for hardware deployment

Check your understanding​

Try to recall each answer before expanding it.

Q1. What do you know about MuJoCo (Multi-Joint dynamics with Contact)?

acquired by DeepMind in 2021, made free; gold standard for contact physics accuracy; used in OpenAI Gym, all major RL benchmarks

Q2. What do you know about MuJoCo 3.0 (2023)?

10-100× faster than previous versions with parallel simulation; runs 10,000+ environments on a single workstation GPU

Q3. What do you know about Isaac Gym (2021, now deprecated)?

NVIDIA's GPU-parallel RL training environment; 4096+ parallel environments on a single A100; enabled ETH Zurich ANYmal locomotion results

Q4. What do you know about Isaac Lab (2024)?

replaces Isaac Gym; built on Isaac Sim 4.0/Omniverse; better asset management, more realistic contact model, ray-traced rendering for visual RL

Q5. What do you know about Isaac Sim 4.0?

full USD (Universal Scene Description) workflow; Omniverse-based; photorealistic rendering for domain randomization of visual appearance

References​

  • Isaac Lab: A Unified and Modular Reinforcement Learning Framework for Robot Learning — Mittal et al. (2023). arXiv 2301.04195

Next: 7.2 Domain Randomization: The Bridge from Sim to Real →

Part of Module 7: Simulation & Digital Twins.