Module 2 — The Digital Twin (Gazebo & Unity)

Focus: Physics simulation and environment building.

You should never debug a balance controller on a real $16,000 humanoid. If the code is wrong, the robot falls, and you have a repair bill. The professional workflow is simulation first: build a digital twin — a physics-accurate virtual copy of the robot and its world — and prove the software there before it ever touches hardware. This is also how you generate the vast training data embodied AI needs.

What you will learn

• Simulating physics, gravity, and collisions in Gazebo — the open-source workhorse tightly integrated with ROS 2.
• High-fidelity rendering and human-robot interaction in Unity — photorealism and interactive environments.
• Simulating sensors — LiDAR, depth cameras, and IMUs that publish the same ROS 2 topics a real sensor would.

Two simulators, two jobs

Gazebo and Unity are not competitors here — they play different roles.

	Gazebo	Unity
Strength	Accurate rigid-body physics, native ROS 2	Photorealistic rendering, rich interaction
Best for	Control, locomotion, sensor physics	Perception data, human-robot interaction, VR
Physics engine	ODE / Bullet / DART	PhysX
ROS 2 link	ros_gz_bridge (built in)	Unity Robotics Hub / ROS-TCP-Connector

A common pipeline: tune locomotion and control in Gazebo, then drop the robot into a Unity scene to generate photorealistic camera data for training perception models.

The "twin" principle: identical interfaces

The reason simulation works is that the simulated robot exposes the exact same ROS 2 interface as the real one. The same /cmd_vel topic drives it; the same /scan and /camera/image_raw topics come out. Your control and perception nodes cannot tell the difference — which means code that works in sim has a real chance of working on hardware.

        ┌──────────────── identical ROS 2 interface ────────────────┐
        │                                                            │
   /cmd_vel ─►  [ Gazebo digital twin ]  ─► /scan /imu /camera     ≡  /cmd_vel ─► [ Real robot ] ─► /scan /imu /camera

This is also where the famous sim-to-real gap lives: friction, sensor noise, and contact dynamics are never perfectly modeled. Module 3 attacks that gap with photorealistic simulation and domain randomization in NVIDIA Isaac.

How this module fits

Module 1 gave the robot a body (URDF) and a nervous system (ROS 2). This module gives it a world and the physics that make that world real. With a working digital twin, Module 3 can train AI perception and navigation at scale, and Module 4 can safely test an LLM commanding the whole system.

Start with Simulating Physics in Gazebo →.