Isaac Lab

Practical

Isaac Lab is NVIDIA’s open-source, GPU-accelerated framework for robot learning built on Isaac Sim. It enables training reinforcement learning and imitation learning policies at massive scale—4096+ parallel environments on a single GPU—making sophisticated robot training practical without supercomputer access.

Prerequisites

Isaac Sim NVIDIA's photorealistic robotics simulation platform

Reinforcement Learning Training policies through trial and error

Why Isaac Lab Matters

Challenge	Isaac Lab Solution
Training is slow on CPU	GPU-parallel: 4096+ environments simultaneously
Need expensive hardware	Train on single RTX GPU, ~90,000 FPS
Sim-to-real gap	Domain randomization + photorealistic rendering
Complex sensor simulation	Full sensor suite: cameras, LiDAR, IMU, contact
Fragmented tools	Unified framework for RL, IL, motion planning

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                         Isaac Lab                                │
├─────────────────────────────────────────────────────────────────┤
│  ┌──────────────────┐  ┌──────────────────┐  ┌───────────────┐  │
│  │ RL Frameworks    │  │ Imitation        │  │ Motion        │  │
│  │ RSL-RL, SKRL     │  │ Learning         │  │ Planning      │  │
│  │ RL-Games, SB3    │  │ (Mimic)          │  │ (cuRobo)      │  │
│  └────────┬─────────┘  └────────┬─────────┘  └───────┬───────┘  │
│           │                     │                     │          │
│  ┌────────▼─────────────────────▼─────────────────────▼───────┐  │
│  │              Environment API (Manager/Direct)               │  │
│  │  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────┐  │  │
│  │  │ Robots      │  │ Sensors     │  │ Domain              │  │  │
│  │  │ Actuators   │  │ Cameras,IMU │  │ Randomization       │  │  │
│  │  │ Controllers │  │ Contact,Ray │  │ ADR, PBT            │  │  │
│  │  └─────────────┘  └─────────────┘  └─────────────────────┘  │  │
│  └────────────────────────────┬────────────────────────────────┘  │
├───────────────────────────────┼───────────────────────────────────┤
│                          Isaac Sim                                │
│  ┌─────────────┐  ┌─────────────┐  ┌────────────────────────┐    │
│  │ PhysX 5     │  │ RTX Render  │  │ USD Scene Graph        │    │
│  │ (GPU)       │  │ (Tiled)     │  │ (OpenUSD)              │    │
│  └─────────────┘  └─────────────┘  └────────────────────────┘    │
├───────────────────────────────────────────────────────────────────┤
│                     NVIDIA Omniverse / GPU                        │
└───────────────────────────────────────────────────────────────────┘

Getting Started

System Requirements

Component	Minimum	Recommended
GPU	Volta+ (CC 7.0+)	RTX 4090 / A6000
VRAM	16GB	24GB+
RAM	32GB	64GB+
OS	Ubuntu 22.04	Ubuntu 22.04/24.04
Python	3.10 (Sim 4.X)	3.11 (Sim 5.X)

# Install Isaac Sim first
pip install isaacsim==5.1.0 --extra-index-url https://pypi.nvidia.com

# Clone Isaac Lab
git clone https://github.com/isaac-sim/IsaacLab.git
cd IsaacLab

# Create conda environment and install
./isaaclab.sh --install

# Clone with submodules
git clone --recurse-submodules https://github.com/isaac-sim/IsaacLab.git
cd IsaacLab

# Install dependencies
./isaaclab.sh --install

# Verify installation
./isaaclab.sh -p source/standalone/tutorials/00_sim/create_empty.py

Training Your First Policy

Train a quadruped locomotion policy:

# Train ANYmal-D on flat terrain with RSL-RL
python source/standalone/workflows/rsl_rl/train.py \
    --task Isaac-Velocity-Flat-Anymal-D-v0 \
    --num_envs 4096 \
    --headless

This trains a policy at ~90,000 FPS on an RTX A6000.

Environment Workflows

Isaac Lab provides two workflow styles for creating environments:

Manager-Based (Recommended)

Modular, structured approach with separate managers for observations, rewards, and actions:

from omni.isaac.lab.envs import ManagerBasedRLEnv

# Create parallel environments
env = ManagerBasedRLEnv(
    cfg=MyRobotEnvCfg,
    num_envs=4096,
)

# Training loop
obs = env.reset()
for _ in range(1000000):
    actions = policy(obs)
    obs, rewards, dones, infos = env.step(actions)

Direct (Performance-Critical)

Lower-level API for maximum control, similar to Isaac Gym:

from omni.isaac.lab.envs import DirectRLEnv
from omni.isaac.lab.utils import configclass

@configclass
class MyRobotEnvCfg(DirectRLEnvCfg):
    scene: MySceneCfg = MySceneCfg(num_envs=4096)
    observation_space = 48
    action_space = 12
    episode_length_s = 20.0

Supported RL Frameworks

Framework	Features
RSL-RL	JIT/ONNX export, fast training
SKRL	PyTorch + JAX, multi-agent (MAPPO, IPPO)
RL-Games	Vectorized training
Stable-Baselines3	Extensive docs, numpy-based

Pre-Built Environments (30+)

Locomotion

Isaac-Velocity-Flat-Anymal-D-v0 — Quadruped flat terrain
Isaac-Velocity-Rough-Anymal-C-v0 — Quadruped rough terrain
Isaac-Humanoid-v0 — Humanoid walking

Manipulation

Isaac-Lift-Cube-Franka-v0 — Pick and lift with Franka
Isaac-Stack-Cube-Franka-v0 — Stack cubes

Dexterous (DexSuite v2.3+)

Isaac-Dexsuite-Kuka-Allegro-Lift-v0 — Dexterous lifting
Isaac-Dexsuite-Kuka-Allegro-Reorientation-v0 — Object reorientation

Supported Robots

Quadrupeds: ANYmal-B/C/D, Unitree A1/Go1/Go2, Boston Dynamics Spot

Humanoids: Unitree H1, G1

Manipulators: Franka Panda, Kuka arms, Universal Robots

Hands: Allegro Hand, Unitree three-finger, Inspire five-finger

Policy Export and Deployment

Export trained policies for deployment:

from isaaclab_rl.rsl_rl import export_policy_as_jit

# Export to TorchScript for deployment
export_policy_as_jit(agent, path="policy.pt")

Deploy on Jetson Orin or other edge devices via Isaac ROS.

Performance Metrics

Metric	Value
State-based training	1.6M+ FPS (multi-GPU)
Vision-based training	60K+ FPS (multi-GPU)
Locomotion training	~90,000 FPS (RTX A6000)
Parallel environments	4096+ (single GPU)

Isaac Gym Migration

Migrating from Isaac Gym? Key differences:

Aspect	Isaac Gym	Isaac Lab
Config format	YAML	Python configclass
Scene creation	Manual loop	Cloner API (automatic)
Rendering	Basic	RTX photorealistic
ROS support	No	Yes (ROS 2 bridge)
Soft body physics	No	Yes (PhysX 5)

Isaac Sim Parent simulation platform

Isaac ROS GPU-accelerated ROS 2 packages

cuMotion GPU-accelerated motion planning

Reinforcement Learning Core training paradigm

Sources

Isaac Lab Documentation — Official docs, tutorials, API reference
Isaac Lab GitHub — Source code, BSD-3-Clause license
Isaac Lab arXiv Paper — Technical details on GPU-accelerated learning
Isaac Lab 2.3 Release — DexSuite, ADR, PBT features
Migrating from IsaacGymEnvs — Migration guide