Simulation-based training’s importance in robotics is analogous to internet data’s importance for large language models — simulators can generate robot motion data at near-zero cost and unlimited speed, while real robot data collection is costly, slow, and poses safety risks. A typical robot skill (like folding a towel) may require tens of thousands of demonstrations to learn effectively — impractical in the real world but achievable in hours in simulation.
## Major Simulation Platforms
**NVIDIA Isaac Sim (Isaac Lab)**: physics simulation platform based on NVIDIA Omniverse; supports GPU parallel simulation (single A100 can run thousands of simulation environments in parallel); supports photorealistic ray-traced rendering (improving visual realism, reducing Sim-to-Real visual gap). Deep ROS 2 integration; supports Unitree, Boston Dynamics, and other major robot model imports. Updated to Isaac Lab in 2024 with extensive embodied intelligence training workflow optimizations.
**DeepMind MuJoCo**: one of the de facto physics simulation standards; accurate physics calculations, fast, free to use. Widely used in reinforcement learning communities (many OpenAI Gym environments are MuJoCo-based). Now open-sourced by DeepMind.
**Genesis (PKU open-source)**: open-source physics simulation platform released by Peking University in 2024; using Position-based Dynamics and GPU acceleration; claimed to be 430,000x faster than traditional physics engines — generating significant attention.
## Core Sim-to-Real Challenges
**Reality gap**: policies learned in simulation fail on real robots due to inaccurate physical parameters (real friction coefficients, motor dynamics, sensor noise not matching simulation parameters) and visual domain gaps (simulation rendering vs. real images differ in lighting, texture, and shadows).
**Domain Randomization (DR)**: randomizing simulation physical parameters (masses, friction coefficients, force perturbations) and visual parameters (lighting, colors, backgrounds) during training, making learned policies more solid to parameter changes and improving transfer ability.
See [Embodied Intelligence and AI](https://sunqi.org/embodied-intelligence-ai-en/) and [MuJoCo official site](https://mujoco.org/).
