Isaac Lab is NVIDIA’s open-source framework for training robot control policies, mainly with reinforcement learning and imitation learning, on top of the Isaac Sim simulator. It runs thousands of simulated robots in parallel on one GPU, so a policy that would take months of real-world trial and error can train in hours. It replaces Isaac Gym, which NVIDIA now lists as deprecated (NVIDIA Isaac Gym page). Isaac Lab itself is free under the BSD-3-Clause license; Isaac Sim, which it needs, has its own NVIDIA license.
Isaac Lab vs Isaac Sim vs Isaac Gym
- Isaac Sim is the simulator: physics, sensors, rendering and robot models.
- Isaac Lab is the learning layer on top: task definitions, rewards, observation pipelines, parallel environments and scripts that connect to RL libraries.
- Isaac Gym was the older standalone RL simulator. NVIDIA says it can still be downloaded but is no longer supported, and the Isaac Lab docs include migration guides from IsaacGymEnvs and OmniIsaacGymEnvs.
If you’re choosing between simulators in general, our robot simulator comparison (Korean) compares Isaac Sim with other robot simulators.
Which version to install (October 2026)
Two release lines exist on GitHub releases:
- Isaac Lab 2.3.2 with Isaac Sim 5.1: the stable line on the main branch, using Python 3.11. Start here if you’re learning or following tutorials.
- Isaac Lab 3.0.0 early access: released September 16, 2026, requiring Isaac Sim 6.1 and Python 3.12. It adds a second physics backend (Newton alongside PhysX), swappable renderers and a single isaaclab command-line tool. NVIDIA’s release notes target general availability for late October 2026.
One change matters if you move code between them: 3.0 switches quaternion order from WXYZ to XYZW, so rotations written for 2.x will be wrong in 3.0 unless you convert them.
Hardware and OS requirements
From the Isaac Lab installation guide:
- OS: Ubuntu 22.04 or Windows 11, 64-bit. DGX Spark is supported with some features missing.
- GPU: an NVIDIA RTX GPU with at least 16 GB of VRAM; more if you render cameras.
- RAM: 32 GB or more.
- Driver: a recent production driver, 580.65.06 or later on Linux and 580.88 or later on Windows.
There’s no macOS support. If you don’t have a suitable GPU, NVIDIA’s Isaac Launchable runs Isaac Sim and Isaac Lab in a cloud machine you reach through the browser, using an NVIDIA Brev account.
Install options
The installation guide lists several paths. Most people need the first:
- Isaac Sim via pip, Isaac Lab from source. NVIDIA’s recommended route.
- Isaac Sim binaries, Isaac Lab from source. Useful if you already have an Isaac Sim install.
- Both from source. For people modifying Isaac Sim itself.
- Docker. For servers and clusters.
- Isaac Lab pip packages only. For building extensions; it doesn’t include training scripts or examples.
The quickstart for the stable line runs in this order: create a Python 3.11 conda environment, install PyTorch 2.7.0 with CUDA 12.8, install isaacsim 5.1.0 from NVIDIA’s package index, clone the Isaac Lab repository on the main branch, then run the install script with the –install flag.
Your first training run
The docs’ existing RL scripts page uses the Ant walking task as a first example. On Linux, from the Isaac Lab folder:
- RL-Games: ./isaaclab.sh -p scripts/reinforcement_learning/rl_games/train.py –task Isaac-Ant-v0 –headless
- RSL-RL with a Franka arm reaching task: ./isaaclab.sh -p scripts/reinforcement_learning/rsl_rl/train.py –task Isaac-Reach-Franka-v0 –headless
On Windows, use isaaclab.bat instead of ./isaaclab.sh. The –headless flag trains without opening the viewer, which is much faster; drop it when you want to watch.
Isaac Lab connects to four RL libraries out of the box: RSL-RL, skrl, RL-Games and Stable-Baselines3. RSL-RL is the common choice for legged robots; skrl and Stable-Baselines3 are easier if you already know them.
Manager-based or direct tasks
When you write your own task, the task workflow docs offer two styles:
- Manager-based: rewards, observations, events and terminations are separate config-driven pieces you swap in and out. Good for prototyping and teams.
- Direct: one class implements everything. More code, but full control and easier to optimize for speed.
Our article on choosing between direct and manager-based task design goes deeper once you’re past the first run.
Related tools
- Isaac Lab Mimic generates extra imitation-learning demonstrations from a few human ones.
- Isaac Lab-Arena builds benchmarks for evaluating trained policies in simulation.
Policies trained in Isaac Lab still have to cross the sim-to-real gap before they work on hardware; our sim2real explainer (Korean) covers why.