Isaac Lab: The Unified GPU-Parallel RL Framework for Robot Learning

What is Isaac Lab?

Isaac Lab is an open-source, BSD-3-licensed framework for reinforcement learning on NVIDIA Isaac Sim. Where Isaac Sim is the general-purpose simulator — a large USD-and-PhysX platform usable for everything from digital twins to teleoperation — Isaac Lab is the thinner, RL-specific layer on top that provides environment wrappers, observation and action-space abstractions, reward composition, domain-randomization helpers, and direct integrations with the four most common RL trainers used in robotics research.

Historically, NVIDIA released three related projects: Isaac Gym, the original GPU-parallel RL framework published alongside the 2021 Isaac Gym paper; OmniIsaacGymEnvs, a port of Isaac Gym to Omniverse; and Orbit, an ETH Zurich-led modular framework. In 2024 all three were unified into Isaac Lab, which became the canonical path forward. Teams migrating from Isaac Gym get GPU parallelism and a familiar task-registration API; teams migrating from Orbit get the modular configuration system and the ETH-curated locomotion tasks. Isaac Lab is the current default.

The framework's headline number is throughput. On a single RTX 4090, Isaac Lab steps a Unitree G1 flat-terrain locomotion environment at over 100,000 steps per second across 4,096 parallel instances. A typical locomotion training run — 1 billion steps of PPO with RSL-RL — completes in under four hours on that hardware, versus overnight or multi-day runs on CPU-only simulators.

Installation Quickstart

Isaac Lab sits on top of Isaac Sim, so the first step is installing Isaac Sim (either through the Omniverse Launcher or via pip). Once Isaac Sim is available, clone Isaac Lab and run its bootstrap script:

# Prerequisite: Isaac Sim 4.5+ installed (see /rl-environments/isaac-sim/)
git clone https://github.com/isaac-sim/IsaacLab.git
cd IsaacLab

# One-command install that pulls deps and registers tasks
./isaaclab.sh --install

# Verify installation by listing available tasks
./isaaclab.sh -p source/standalone/environments/list_envs.py

Training a reference locomotion task on the Unitree G1 with RSL-RL takes a single command. Logs stream to TensorBoard, and checkpoints are written to logs/rsl_rl/g1_flat/:

# Train G1 flat-terrain velocity tracking with RSL-RL PPO
./isaaclab.sh -p source/standalone/workflows/rsl_rl/train.py \
  --task Isaac-Velocity-Flat-G1-v0 \
  --num_envs 4096 --headless --max_iterations 1500

Swapping RL trainers is a flag change — the same task can be trained with RL-Games, SKRL, or Stable-Baselines3 via the equivalent workflows/ subdirectory.

Supported Robots and Tasks

Isaac Lab inherits the full Isaac Sim robot library, with pre-configured environments for the most commonly benchmarked platforms. Locomotion is the deepest area: Unitree G1, H1, H1-2, and Go2, Boston Dynamics Spot, ANYbotics ANYmal C and D, and Agility Digit all have flat-terrain and rough-terrain variants with domain-randomized velocity tracking rewards derived from the original Legged Gym paper. The tasks follow a consistent command-tracking formulation that makes cross-robot comparisons straightforward.

For manipulation, Isaac Lab ships with Franka Research 3 reach, pick-and-place, and cube stacking, as well as Universal Robots UR10e tasks. Dexterous manipulation is represented by Allegro and Shadow Hand in-hand cube re-orientation, with sub-millimeter contact simulation from GPU PhysX. Humanoid whole-body control tasks include full-body tracking of reference motions on H1 and G1, closely mirroring the published GR00T and H1 walking papers.

The task list is growing fast: the Isaac Lab repository tracks contributions from universities and labs worldwide, and the 40+ task count from 2025 is expected to double through 2026 with the addition of bimanual manipulation, mobile manipulation, and VLA-evaluation suites.

Benchmarks on Isaac Lab

Isaac Lab provides the canonical open-source implementations of several widely cited baselines. The legged-locomotion suite reproduces the ETH Zurich RSL-RL results from the original Legged Gym paper (Rudin et al., CoRL 2021), which achieved blind rough-terrain traversal on ANYmal. The ORBIT manipulation benchmark, introduced with the Orbit paper (Mittal et al., RA-L 2023), provides 15 standardized manipulation tasks with PPO and SAC baselines. The H1 whole-body tracking benchmark, contributed by NVIDIA Research and Carnegie Mellon, is the reference for humanoid motion imitation.

For humanoid and dexterous research, Isaac Lab's pre-built tasks are the de facto comparison point in every 2025 locomotion paper: published numbers on Unitree G1 velocity tracking, H1 walking, and the ANYmal parkour policy all come from Isaac Lab configurations. Our sim2real digital-twin deliveries use Isaac Lab as the default training stack for legged-robot clients.

Pros and Cons

Strengths. Best-in-class GPU parallelism for RL, BSD-3 permissive license, four interchangeable RL trainers, a large library of reference tasks that reproduce published results, active upstream development from NVIDIA and ETH Zurich, direct access to Isaac Sim's full asset library, and a configuration system that is genuinely modular (swap the robot, swap the terrain, swap the reward — all through config rather than code edits).

Weaknesses. Inherits Isaac Sim's hardware and install requirements (NVIDIA RTX GPU, 40+ GB footprint). The config system has a learning curve — teams coming from Isaac Gym need to re-learn the Orbit-style class hierarchy. Documentation depth is uneven across tasks; locomotion is well-covered, but manipulation and dexterous docs lag. Some Isaac Gym tasks are not yet fully ported, and the version matrix between Isaac Sim, Isaac Lab, and the RL trainers requires careful pinning.

When to Pick Isaac Lab

Choose Isaac Lab when you want GPU-parallel RL on Isaac Sim and do not want to write your own environment plumbing. It is the right default for humanoid locomotion, quadruped traversability, contact-rich manipulation at scale, and any workflow where reproducing a published paper is a first-class requirement. If your lab has standardized on NVIDIA hardware and ROS 2, Isaac Lab removes essentially all integration friction.

Pick MuJoCo instead when CPU-only iteration, installation simplicity, or differentiable dynamics matters more than raw throughput. Pick Isaac Sim directly when you need digital-twin, perception, or non-RL robotics workflows. Pick Robosuite for standardized manipulation benchmarks. See our MuJoCo vs Isaac Sim 2026 guide for the broader simulator comparison.

Get a Custom Isaac Lab Environment

SVRC builds Isaac Lab environments for humanoid and quadruped teams: custom locomotion tasks on your robot, domain-randomized rough-terrain training, dexterous manipulation suites, and sim2real pipelines validated against hardware in our Mountain View lab. Every delivery ships with RSL-RL configs, reproducible seeds, and a TensorBoard log bundle from a reference training run.

Related Links

• RL Environments hub — compare 8 major simulators.
• NVIDIA Isaac Sim — the underlying photorealistic simulator.
• MuJoCo — CPU-fast alternative for contact-rich research.
• MuJoCo vs Isaac Sim 2026 — head-to-head comparison.
• Compatible hardware in the store — Unitree G1, ANYmal, Spot.
• Custom teleoperation datasets to seed imitation learning.