Isaac Sim vs Isaac Lab: Understanding the Stack

ROSCon India 2025 Workshop Prep - Part 1 of 4

Isaac Sim vs Isaac Lab: What’s the Difference?

Before diving into Go2-W training, let’s clarify the NVIDIA stack—because the naming can be confusing:

┌─────────────────────────────────────────────────────────────┐
│                    NVIDIA ROBOTICS STACK                    │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  ┌─────────────────────────────────────────────────────┐   │
│  │              robot_lab Extension                     │   │
│  │         (Go2-W environments - what we need!)        │   │
│  └─────────────────────────────────────────────────────┘   │
│                          ↓                                  │
│  ┌─────────────────────────────────────────────────────┐   │
│  │                 ISAAC LAB                            │   │
│  │     Reinforcement Learning Framework                 │   │
│  │     • RL training environments                       │   │
│  │     • PPO, SAC algorithms (via RSL-RL)              │   │
│  │     • Domain randomization                           │   │
│  │     • Curriculum learning                            │   │
│  └─────────────────────────────────────────────────────┘   │
│                          ↓                                  │
│  ┌─────────────────────────────────────────────────────┐   │
│  │                 ISAAC SIM                            │   │
│  │     Physics Simulation Platform                      │   │
│  │     • PhysX 5 rigid body dynamics                   │   │
│  │     • USD scene format                               │   │
│  │     • RTX rendering                                  │   │
│  │     • ROS 2 bridge                                   │   │
│  └─────────────────────────────────────────────────────┘   │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Component	Purpose	When to Use
Isaac Sim	Physics simulation, visualization	Loading robots, testing scenes, ROS 2 integration
Isaac Lab	RL training framework	Training locomotion policies, domain randomization
robot_lab	Go2-W specific environments	Training the Go2-W (not included in stock Isaac Lab)

We’re Using Isaac Lab

This workshop focuses on reinforcement learning, which means we need Isaac Lab (not just Isaac Sim). Isaac Lab includes Isaac Sim as its foundation—when you run Isaac Lab, you get the full simulation capabilities plus the RL training framework.

---

Launching Isaac Lab

We’ve already set up a convenient launcher script in an earlier post: Setting Up Your Physical AI Laboratory.

Quick Launch

# From anywhere (using the symlink we created)
~/auto_start.sh

# Select Option 8: Isaac Lab: Start/Build Container

What Option 8 Does

1. First run: Builds the Isaac Lab Docker container (~10-15 min)
2. Subsequent runs: Starts the existing container
3. Enters an interactive shell inside the container
4. Mounts your Isaac Lab repository for persistent changes

┌─────────────────────────────────────────────────────────────┐
│              AUTO_START.SH MENU (Option 8)                   │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  Isaac Lab Container includes:                              │
│  ─────────────────────────────                              │
│  • Isaac Sim 5.1.0 (simulation engine)                     │
│  • Isaac Lab framework (RL training)                       │
│  • RSL-RL (PPO implementation)                             │
│  • PyTorch with CUDA                                       │
│  • All dependencies pre-configured                         │
│                                                             │
│  Your robot_lab extension installs INTO this container     │
│                                                             │
└─────────────────────────────────────────────────────────────┘

Why Use the Container?

The Isaac Lab Docker container ensures all dependencies are correctly versioned and configured. No “works on my machine” issues!

---

Our Mission: Train the Go2-W

This 4-part series focuses exclusively on training the Unitree Go2-W—the wheeled variant of the Go2 quadruped. The Go2-W combines legged locomotion with wheeled mobility, creating a 16-DoF hybrid platform that requires specialized simulation setup.

Why Go2-W Specifically?

Simple—I own one, and it’s the only robot that can conquer the stairs at home!

• Ground clearance: The 7-inch wheels provide significantly more clearance than the standard Go2
• Uneven stairs: My home has stairs that are uneven and taller than what the regular Go2 can handle
• Hybrid advantage: When wheels alone can’t cut it, the legs take over—best of both worlds

Training a policy for this platform is fundamentally different from standard quadrupeds, and that’s exactly why it’s interesting.

┌─────────────────────────────────────────────────────────────┐
│                    GO2-W HYBRID PLATFORM                     │
│                                                             │
│     ┌─────────────────────────────────────────┐             │
│     │           LEGS (12 DoF)                 │             │
│     │  Hip → Thigh → Calf (×4 legs)          │             │
│     └─────────────────────────────────────────┘             │
│                        +                                    │
│     ┌─────────────────────────────────────────┐             │
│     │          WHEELS (4 DoF)                 │             │
│     │  7-inch pneumatic tires (×4)           │             │
│     └─────────────────────────────────────────┘             │
│                        =                                    │
│     ┌─────────────────────────────────────────┐             │
│     │     16-DoF HYBRID LOCOMOTION            │             │
│     │  Roll on flat, walk over obstacles     │             │
│     └─────────────────────────────────────────┘             │
└─────────────────────────────────────────────────────────────┘

Series Navigation

• Part 1: Isaac Sim vs Isaac Lab ← You are here
• Part 2: Go2-W ROS 2 Interface
• Part 3: Training Go2-W Locomotion
• Part 4: Go2-W Sim-to-Real

---

Go2-W Specifications

Specification	Value
Dimensions	70cm × 43cm × 50cm
Weight	~18kg (including battery)
Leg Joints	12 DoF (Hip, Thigh, Calf × 4)
Wheel Motors	4 DoF (in-wheel motors)
Total DoF	16
Tires	7-inch pneumatic
Max Wheeled Speed	2.5 m/s
Max Climb Angle	35 degrees
Obstacle Height	Up to 70cm (walking mode)

Joint Structure (16 DoF)

Go2-W Joint Names:
──────────────────────────────────────────────────
Front Left (FL):
  FL_hip_joint    → FL_thigh_joint → FL_calf_joint → FL_wheel_joint

Front Right (FR):
  FR_hip_joint    → FR_thigh_joint → FR_calf_joint → FR_wheel_joint

Rear Left (RL):
  RL_hip_joint    → RL_thigh_joint → RL_calf_joint → RL_wheel_joint

Rear Right (RR):
  RR_hip_joint    → RR_thigh_joint → RR_calf_joint → RR_wheel_joint
──────────────────────────────────────────────────
Legs (12):  Position/Torque control
Wheels (4): Velocity control (continuous rotation)

---

The Simulation Asset Challenge

Critical Finding

A direct, official USD for the Go2-W does not exist in NVIDIA’s built-in assets or Unitree’s primary release channels.

You cannot simply “find” the Go2-W—you must construct it or use the robot_lab extension.

Two Paths to Go2-W in Isaac Sim

Path	Complexity	Use Case
Option A: robot_lab Extension	Easy ✅	Ready-to-train RL environments
Option B: URDF Import	Advanced	Custom USD asset creation

---

Option A: robot_lab Extension (Recommended)

The robot_lab extension by fan-ziqi provides pre-configured Go2-W environments for Isaac Lab.

Installation

# Clone the extension
git clone https://github.com/fan-ziqi/robot_lab.git
cd robot_lab

# Install into Isaac Lab
python -m pip install -e source/robot_lab

Verify Go2-W Environment

# List available environments
./isaaclab.sh -p scripts/environments/list_envs.py | grep -i go2w

Expected output:

RobotLab-Isaac-Velocity-Flat-Unitree-Go2W-v0
RobotLab-Isaac-Velocity-Rough-Unitree-Go2W-v0

Quick Test

# Train Go2-W (quick test)
./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py \
    --task=RobotLab-Isaac-Velocity-Rough-Unitree-Go2W-v0 \
    --num_envs=64 \
    --max_iterations=100

# Play trained policy
./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/play.py \
    --task=RobotLab-Isaac-Velocity-Rough-Unitree-Go2W-v0

Compatibility: Isaac Sim 4.5/5.0/5.1 + Isaac Lab 2.3.0

---

Option B: Custom URDF Import

If you need a custom Go2-W USD asset (for custom sensors, modifications, etc.), import from the official URDF.

Source Repositories

Repository	URL	Content	Path
unitree_ros	GitHub	URDF, Meshes	robots/go2w_description/
unitree_mujoco	GitHub	Physics Params	unitree_robots/go2w/scene.xml

Step 1: Get the URDF

# Clone unitree_ros
git clone https://github.com/unitreerobotics/unitree_ros.git

# Go2-W files are at:
# unitree_ros/robots/go2w_description/
#   ├── urdf/go2w_description.urdf
#   ├── xacro/
#   └── meshes/

Step 2: Fix Package Paths

The URDF uses ROS package paths that Isaac Sim cannot resolve:

<!-- Before: ROS package path -->
<mesh filename="package://go2w_description/meshes/base_link.dae"/>

<!-- After: Relative path -->
<mesh filename="../meshes/base_link.dae"/>

Step 3: Import to Isaac Sim

1. Open Isaac Sim
2. Navigate to: Isaac Utils > Workflows > URDF Importer
3. Select go2w_description.urdf
4. Configure import settings:

Setting	Value	Reason
Fix Base Link	❌ Unchecked	Go2-W is mobile
Leg Joint Drive Type	Position or Effort	RL outputs positions/torques
Wheel Joint Drive Type	Velocity	Wheels need continuous rotation
Self-Collision	❌ Disabled	Performance optimization
Create Physics Scene	✅ Checked	Required for simulation

Step 4: Fix the “Pink Robot” Issue

Known Issue: GitHub Issue #131

After import, the robot may appear pink or white (untextured).

Cause: Incompatibility between Unitree’s Collada materials and Omniverse MDL.

Manual Fix:

1. Expand Stage: Go2W > base_link > Visuals
2. Select mesh prims
3. Apply OmniPBR materials:

Material	Albedo (Hex)	Roughness	Metallic	Apply To
Rubber_Black	#1A1A1A	0.8	0.0	Wheel meshes
Unitree_Grey	#C0C0C0	0.4	0.6	Body and legs

4. Save as go2w_conditioned.usd

---

Physics Configuration

Leg Joints (Revolute)

Parameter	Value	Notes
Stiffness	0 or low	Let RL policy be the controller
Damping	1.0–5.0 Ns/m	Motor back-EMF and friction

Joint Limits

Joint	Lower Limit	Upper Limit
Hip	-1.04 rad	+1.04 rad
Thigh	-1.57 rad	+3.49 rad
Calf	-2.72 rad	-0.83 rad
Wheel	−∞	+∞

Wheel Joints (Continuous)

Parameter	Value	Notes
Drive Type	Velocity	Continuous rotation
Friction	Low	Efficient rolling
Damping	Minimal	Prevent “sticky” behavior

Tire Physics Materials

Parameter	Value	Reason
Static Friction	1.0	Prevent sideways sliding
Dynamic Friction	0.8	Traction during acceleration
Restitution	0.05	Pneumatic tires absorb impact

Complete Physics Material Table

Material	Static	Dynamic	Restitution	Use
Tire Rubber	1.0	0.8	0.05	Wheels
Aluminum	0.5	0.3	0.1	Chassis/Legs
Concrete	0.8	0.7	0.1	Training Ground
Ice	0.1	0.05	0.1	Domain Randomization

---

Inertial Property Validation

Critical Step

The Go2-W weighs 18kg—3kg more than the standard Go2. Using wrong inertia values will cause unstable simulation.

Validation Process:

1. Select base_link prim
2. Check Physics > Mass API
3. Verify trunk mass: ~6-8kg (total should sum to ~18kg)
4. If values are wrong, reference unitree_mujoco/unitree_robots/go2w/scene.xml

---

Troubleshooting

Problem: Robot Appears Pink/White

Cause: Material path issues during URDF import

Fix: Manually apply OmniPBR materials (see Step 4 above)

Problem: Simulation Explodes

Causes: - Invalid inertia tensor (too small or non-positive definite) - Incorrect collision meshes

Fixes: - Re-check mass properties in Physics API - Use Convex Hulls instead of Triangle Meshes for collisions

Problem: Wheels Drag Instead of Roll

Causes: - Wheel joint set to Position drive (should be Velocity) - Zero friction on wheel physics material - Joint type not set to Continuous

Fix: Ensure wheel joints use Velocity drive mode

---

Quick Reference Card

┌─────────────────────────────────────────────────────────────┐
│              GO2-W ISAAC SIM QUICK REFERENCE                 │
├─────────────────────────────────────────────────────────────┤
│                                                             │
│  RECOMMENDED: Use robot_lab Extension                       │
│  ─────────────────────────────────────                     │
│  git clone https://github.com/fan-ziqi/robot_lab.git       │
│  python -m pip install -e source/robot_lab                  │
│                                                             │
│  Task: RobotLab-Isaac-Velocity-Rough-Unitree-Go2W-v0       │
│                                                             │
│  ───────────────────────────────────────────────────────── │
│                                                             │
│  MANUAL IMPORT: URDF → USD                                  │
│  ─────────────────────────────                             │
│  URDF: unitree_ros/robots/go2w_description/urdf/           │
│  Physics: unitree_mujoco/unitree_robots/go2w/scene.xml     │
│                                                             │
│  Import Settings:                                           │
│    • Fix Base Link: OFF                                    │
│    • Leg Joints: Position/Effort                           │
│    • Wheel Joints: VELOCITY (important!)                   │
│                                                             │
│  ───────────────────────────────────────────────────────── │
│                                                             │
│  GO2-W SPECS                                                │
│  ───────────                                               │
│  Total Joints: 16 (12 leg + 4 wheel)                       │
│  Weight: 18kg                                              │
│  Max Speed: 2.5 m/s (wheeled)                              │
│                                                             │
└─────────────────────────────────────────────────────────────┘

---

Screenshots to Capture

When setting up Go2-W, capture these for reference:

• robot_lab extension installed successfully
• Go2-W environments listed in terminal
• Go2-W spawned in Isaac Sim (if using custom USD)
• Physics materials applied correctly
• Stage hierarchy showing 16 joints

---

Workshop Questions

Questions to Ask

1. Is Go2-W planned for official Isaac Lab assets?
2. Best practices for tire-terrain friction in PhysX?
3. How to validate inertia tensor against real hardware?
4. Any known issues with robot_lab + Isaac Lab 2.3.0?

---

What’s Next

With Go2-W set up in Isaac Sim, we can now connect it to ROS 2:

→ Part 2: Go2-W ROS 2 Interface - 16-joint control via ROS 2

---

Sources

• robot_lab Extension - Go2-W Isaac Lab environments
• unitree_ros - Official URDF source
• unitree_mujoco - Physics parameters
• Isaac Sim Documentation
• URDF Importer Guide