NVIDIA Isaac Lab
Last reviewed
Apr 30, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 3,571 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 30, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 3,571 words
Add missing citations, update stale details, or suggest a clearer explanation.
NVIDIA Isaac Lab is an open-source, GPU-accelerated unified framework for robot learning, built on top of NVIDIA Isaac Sim. It supports reinforcement learning, imitation learning, and motion planning at scale, with thousands of parallel environments running on a single GPU. The framework was originally announced as Orbit in early 2023 by the NVIDIA Seattle Robotics Lab, and was rebranded as Isaac Lab in March 2024 alongside the unveiling of Project GR00T at NVIDIA's GTC conference.
Isaac Lab is licensed under the BSD-3-Clause license and is hosted on GitHub at isaac-sim/IsaacLab. It is the consolidated successor to several earlier NVIDIA simulation efforts, including the original Isaac Gym, the IsaacGymEnvs benchmark suite, the OmniIsaacGymEnvs wrapper for Isaac Sim, and the Orbit framework. Robotics groups at ETH Zurich, ANYbotics, Boston Dynamics, Toyota Research Institute, Apptronik, 1X, Galbot, Fourier, Unitree, and many universities use it to train locomotion and manipulation policies that transfer onto physical hardware.
The core idea behind Isaac Lab is straightforward: keep both the physics simulation and the policy network on the GPU, so the slow CPU-to-GPU data transfer that bottlenecks classical reinforcement learning loops disappears. PhysX 5 simulates rigid bodies, articulated systems, and deformable objects in batched form. Observations are read directly from PhysX buffers as PyTorch tensors. The policy gradient step happens in the same memory space. The result is that thousands of parallel agents can step a complex scene at hundreds of thousands of physics steps per second, on a single workstation card.
Isaac Lab inherits this all-on-GPU design from its predecessor Isaac Gym (Makoviychuk et al., 2021), but it adds the more general scene authoring of NVIDIA Omniverse and Universal Scene Description (USD), photoreal RTX rendering, sensor models for cameras, LiDAR, IMUs and force-torque, and a manager-based modular API for environment construction. The framework targets researchers building reinforcement learning policies, engineers collecting imitation learning datasets, and practitioners testing motion planners under realistic contact dynamics.
A 2025 paper led by Mayank Mittal with 104 co-authors, titled "Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning," formalised the design and described its use across whole-body humanoid control, quadruped mobility, contact-rich assembly, and dexterous manipulation.
Isaac Lab is the third generation of NVIDIA's GPU-native robot simulation work. The lineage runs from Isaac Gym (a research preview that proved the all-on-GPU concept) to Orbit (a USD-based framework on top of Isaac Sim) to Isaac Lab (the current consolidated platform under the Isaac brand).
| Year | Project | Notes |
|---|---|---|
| 2021 | Isaac Gym Preview | First released as a standalone preview. PyTorch-based GPU physics and rendering, used for ANYmal blind locomotion (Rudin et al.), Shadow Hand cube manipulation, and quadruped tasks. NeurIPS Datasets and Benchmarks paper by Makoviychuk et al. |
| 2022 | OmniIsaacGymEnvs | Wrapper that ported the IsaacGymEnvs task suite to run inside Isaac Sim on Omniverse, allowing USD scenes and RTX rendering. |
| Jan 2023 | Orbit (arXiv 2301.04195) | Mayank Mittal et al. release a unified framework for interactive robot learning on top of Isaac Sim, with photoreal scenes, fixed-arm and mobile manipulators, and benchmark tasks for cabinet opening, cloth folding and room reorganisation. Hosted at isaac-orbit.github.io. |
| 2023 | Orbit published | Paper appears in IEEE Robotics and Automation Letters, vol. 8, no. 6, pp. 3740 to 3747, DOI 10.1109/LRA.2023.3270034. |
| Mar 2024 | Isaac Lab announced | At GTC 2024 (18 March 2024) NVIDIA unveiled Project GR00T and announced the rebrand of Orbit as Isaac Lab. The new framework absorbed Isaac Gym and OmniIsaacGymEnvs functionality and became the official robot-learning surface on top of Isaac Sim. |
| Jul 2024 | Isaac Lab 1.0 | First general-availability release. Added an LLM-to-reward function reference example and flexible task-design workflows. |
| Nov 2024 | Isaac Lab 1.2 | General availability. Earlier RL frameworks (IsaacGymEnvs, OmniIsaacGymEnvs, Orbit) declared deprecated in favour of Isaac Lab. |
| 2025 | Isaac Lab 2.0 | Major rewrite aligned with Isaac Sim 4.5. Up to 1.2x faster tiled rendering. Introduced the Isaac Lab Mimic extension for automatic trajectory augmentation in imitation learning. |
| 2025 | Isaac Lab 2.1, 2.2 | 2.1 added Apple Vision Pro teleoperation, bi-manual teleop, and USD-level randomisation. 2.2 (on Isaac Sim 5.0) added FORGE and AutoMate contact-rich tasks and GR1 humanoid mimic environments. |
| 2025 | Isaac Lab 2.3 | Built on Isaac Sim 5.1. Added DexSuite dexterous manipulation environments, surface-gripper workflow, SkillGen integration with cuRobo, G1 humanoid loco-manipulation, and expanded teleoperation devices (Vive, Manus glove). |
| Nov 2025 | Isaac Lab paper | arXiv preprint 2511.04831 by Mittal et al. with 104 co-authors documents the framework's architecture and points to the upcoming Newton differentiable physics engine. |
Isaac Lab sits on top of Isaac Sim, which itself is built on Omniverse Kit and uses Pixar's Universal Scene Description format for scenes and robot assets. The physics engine is NVIDIA PhysX 5, configured to run thousands of independent worlds in parallel, each with its own contacts and articulations. The renderer is RTX, which supports both fast tiled rendering for camera observations and photoreal path tracing for synthetic data.
The code base is organised around a small number of long-lived concepts:
Observations and actions are batched across environments with a leading batch dimension, so a typical Isaac Lab loop processes a tensor of shape (num_envs, obs_dim) per step and emits one of shape (num_envs, act_dim). The framework supports both single-agent and multi-agent training and supports cross-embodiment training (for example, the same policy network learning across several quadruped morphologies).
Isaac Lab is built to scale beyond a single machine. PyTorch's distributed framework drives multi-GPU and multi-node training, with NCCL handling collective communication between workers. Industrial users routinely train on DGX-class servers; the announcement of NVIDIA DGX Spark and the Blackwell generation has been bundled with updated Isaac Lab support.
Isaac Lab does not ship its own RL algorithm implementations. Instead, it provides wrappers that adapt its ManagerBasedRLEnv interface to the conventions of several established libraries. The choice of library is left to the user.
| Library | Origin | Style | Notes |
|---|---|---|---|
| RSL-RL | Robotic Systems Lab, ETH Zurich | PyTorch, vectorised, on-policy (PPO) | Closely tied to ETH's quadruped locomotion work; default choice for ANYmal, Spot, Unitree examples. Updated to v2.3.0+ in Isaac Lab 2.1.1. |
| RL Games | Denys Makoviychuk | PyTorch, GPU-vectorised, supports PPO, A2C, SAC | Original library used by the Isaac Gym Preview. Still preferred for Shadow Hand cube and many manipulation tasks. |
| skrl | Antonio Serrano-Munoz | PyTorch and JAX, modular | Most flexible algorithm coverage and supports both backends. |
| Stable-Baselines3 | DLR-RM | PyTorch, CPU-friendly VecEnv API | Provides a familiar interface for users coming from CPU-based gym workflows; less efficient at high num_envs. |
| Robomimic | Stanford / Toyota Research | PyTorch, imitation learning | Used for behaviour cloning and offline RL on demonstrations collected through Isaac Lab teleoperation. |
| Isaac Lab Mimic | NVIDIA | Demonstration augmentation | Generates additional trajectories from a small set of human teleop demos. Apache 2.0 licensed extension introduced in Isaac Lab 2.1. |
The environment classes follow the Gym and Gymnasium API conventions, with the standard reset, step, observation_space, and action_space semantics, but expanded to handle batched trajectory data and per-environment episode bookkeeping.
A simulator that does not transfer is a video game. Isaac Lab leans heavily on sim-to-real transfer tooling so that policies trained in the framework can be deployed on real robots without extensive re-tuning.
Domain randomisation is implemented through event terms in the manager-based environment. The framework can randomise mass, inertia, joint friction, joint damping, joint stiffness, contact friction, payload, controller delays, and motor noise on every reset or every several resets. Visual randomisation in 2.1 added USD-level texture, colour, and scale randomisation, which makes the rendered images vary enough that vision policies do not overfit to a single fixed scene.
Sensor noise models cover camera depth noise, RGB compression, IMU bias drift, and contact-force noise. The RayCaster sensor approximates a LiDAR by sampling per-environment ray directions and intersecting with the USD scene; this is fast enough to use as an observation at training time. Force-torque sensors at end effectors and feet are common for contact-rich manipulation and for legged locomotion respectively.
For visual sim-to-real, Isaac Lab integrates with NVIDIA Cosmos, the world foundation model family, to augment rendered images with photoreal video and to fill in the visual gap between synthetic frames and real camera footage. The 2025 NVIDIA Isaac GR00T Blueprint pipeline used Cosmos Transfer to multiply a small set of teleoperated demonstrations into 780,000 synthetic trajectories in 11 hours, equivalent to roughly 6,500 hours of human demonstration data.
Isaac Lab ships with more than 16 pre-integrated robots and over 30 ready-to-train environments. The list keeps growing release by release; below is the picture as of the 2.3 release in late 2025.
| Category | Robots | Common tasks |
|---|---|---|
| Quadrupeds | ANYmal B, ANYmal C, ANYmal D (ANYbotics); Spot (Boston Dynamics); A1, Go1, Go2 (Unitree); MIT Cheetah-derived models | Velocity tracking on flat terrain, rough-terrain locomotion, stand and recover, blind locomotion, perceptive locomotion |
| Bipeds | Cassie (Agility Robotics origin), Digit (Agility Robotics) | Walking, standing, balance |
| Humanoids | Unitree H1, Unitree G1, Fourier GR-1, Galbot, Agibot platforms; G1 with Inspire dexterous hand | Whole-body locomotion, loco-manipulation, teleoperated mimic, dexterous tasks |
| Single-arm manipulators | Franka Emika Panda, Universal Robots UR5/UR10/UR10e, KUKA iiwa, Open Arm | Reach, lift, stack, drawer opening, peg-in-hole assembly, gear meshing (FORGE, AutoMate) |
| Dual-arm manipulators | Bimanual UR, bimanual Open Arm | Cloth folding, room reorganisation, hand-over |
| Dexterous hands | Shadow Hand, Allegro Hand, Inspire hand on G1 | In-hand cube reorientation, fingertip manipulation, DexSuite suite (2.3) |
| Mobile manipulators | KUKA + Allegro, Franka on mobile base, dedicated parallel pick-and-place platform | Pick-and-place, mobile reach |
| Aerial | Crazyflie-style quadcopter | Hover and trajectory tracking |
The framework is also a foundation for NVIDIA's own Isaac GR00T workflows, which target humanoid manipulation. GR00T-Mimic generates synthetic trajectories inside Isaac Lab, GR00T-Gen randomises scene backgrounds and lighting, and GR00T-Dexterity, GR00T-Mobility, GR00T-Control and GR00T-Perception each consume Isaac Lab environments for specific subtasks of humanoid behaviour.
Most robots ship with an MJCF or URDF source and a curated USD asset. The MJCF and URDF importers were updated in v2.0 and the URDF importer gained merge-joints support in v2.3.1. Users can drop a new robot into Isaac Lab by writing a configuration class that describes its joints, actuators, and default initial state.
Quadruped locomotion. ETH Zurich's Robotic Systems Lab developed RSL-RL specifically for the Isaac Gym workflow and continues to use Isaac Lab to train ANYmal locomotion policies. The same lab's perceptive locomotion work, in collaboration with ANYbotics, transferred policies trained on tens of thousands of parallel terrain instances to physical ANYmal robots in the field.
Boston Dynamics Spot. Boston Dynamics has used Isaac Lab to train policies that are deployed directly for inference on Spot. The company has spoken publicly about how the GPU-parallel rollouts collapse weeks of CPU-based training into hours.
Project GR00T. NVIDIA's humanoid foundation model effort, announced by Jensen Huang at GTC on 18 March 2024, uses Isaac Lab as its primary policy-training environment. Subsequent releases including GR00T N1 (announced in March 2025) and GR00T N1.6 use Isaac Lab for both reinforcement learning and the imitation learning pipeline. At GTC, Huang demonstrated 1X's humanoid robot performing autonomous domestic tidying using a post-trained policy built on GR00T.
Apptronik, Fourier, Galbot. Several humanoid robotics startups have publicly described training workflows that use Isaac Lab for whole-body and manipulation policy training, with Cosmos-based image augmentation to bridge the visual gap.
University research. The framework underpins many recent results in dexterous manipulation, contact-rich assembly, and cross-embodiment locomotion. The 2025 Mittal et al. paper alone has 105 authors drawn from NVIDIA, ETH Zurich, MIT, Stanford, CMU, Berkeley, and others.
Isaac Lab is one of several robot-learning simulators in active use. Each has different trade-offs around physics fidelity, GPU acceleration, scene authoring, and ROS integration.
| Simulator | Maintainer | Physics | GPU acceleration | Scene format | Strengths | Weaknesses |
|---|---|---|---|---|---|---|
| Isaac Lab | NVIDIA | PhysX 5 | Native, GPU-resident | USD | Thousands of parallel envs; modern API; Cosmos and GR00T integration; broad robot zoo | Tied to NVIDIA hardware; Isaac Sim install heavy; Linux preferred |
| Isaac Gym (preview) | NVIDIA | PhysX | Native, GPU-resident | Custom | Pioneered all-on-GPU RL; very fast | Deprecated end-of-2024; superseded by Isaac Lab |
| MuJoCo | DeepMind / Google | MuJoCo | Via MJX (JAX) on GPU/TPU | XML | Excellent contact accuracy; trusted in research; small footprint | MJX still maturing; classic MuJoCo is CPU only |
| PyBullet | Erwin Coumans / open source | Bullet | CPU only in Python bindings | URDF, SDF | Easy to install; popular for quick prototyping | Slow at scale; limited rendering |
| Gazebo (Ignition) | Open Robotics | ODE, DART, Bullet, others | None | SDF | Tight ROS integration; multi-robot scenarios; widely used in industry | Slower physics; heavier installation |
| Genesis | Genesis Embodied AI | Custom (Taichi-based) | Native | Python-native | Recent (2024); claims very high throughput; pure Python | Young project; smaller asset library |
| Drake | Toyota Research Institute | Drake's own | None at user-facing level | URDF, SDF | Strong dynamics correctness; symbolic reasoning; useful for control research | Steeper learning curve; smaller RL community |
| Webots | Cyberbotics | ODE | None | Custom | Education-friendly; cross-platform | Lower fidelity for research-grade RL |
The practical pattern in 2025 is that Isaac Lab dominates for GPU-parallel reinforcement learning at scale, MuJoCo (often via MJX) holds the trust of fundamental contact-mechanics research, and Gazebo remains the standard whenever the work needs to fit into a ROS pipeline or to integrate many heterogeneous nodes.
The public surface of Isaac Lab is a Python API. Internally, the stack runs roughly as follows:
isaaclab package, which exposes scene, asset, sensor, actuator, manager, and environment classes.omni.isaac.lab and broader Omniverse Kit extensions, which are loaded inside an Isaac Sim runtime.torch.distributed backend with NCCL collectives splits environment shards across workers.Deterministic physics is available where the underlying solver is configured for it, but in most reinforcement-learning workloads small variations in solver iterations and per-environment randomisation make pure determinism less interesting than reproducibility through fixed seeds.
Isaac Lab works with both the standard Isaac Sim install and the pip install isaacsim packaged path introduced in 2.0. Asset libraries (USD scenes, robot meshes, materials) are downloaded from NVIDIA's Nucleus servers and cached locally.
Isaac Lab itself is open source under the BSD-3-Clause license. The isaaclab_mimic extension is Apache 2.0. Both can be cloned, modified, and redistributed with attribution.
Isaac Sim, on which Isaac Lab depends, is more nuanced. The Isaac Sim source code on isaac-sim/IsaacSim is released under Apache 2.0, but building or running it requires the Omniverse Kit SDK and the bundled assets, which carry separate proprietary terms. Internal R&D use is free of charge. Redistributing Isaac Sim with Omniverse Kit, or delivering it as a service to third parties, requires an NVIDIA AI Enterprise licence. In practical terms most academic and individual users treat Isaac Sim as a free download from NVIDIA, and that is the supported path the documentation describes.
Hardware and operating system requirements as of Isaac Lab 2.3 and Isaac Sim 5.1:
| Component | Minimum | Recommended |
|---|---|---|
| Operating system | Ubuntu 22.04 (Linux x64) or Windows 11 (x64) | Ubuntu 22.04 |
| CPU | Modern x64 with AVX2 | Multi-core workstation CPU |
| RAM | 32 GB | 64 GB or more |
| GPU | NVIDIA RTX with at least 8 GB VRAM for small scenes | RTX 30, 40, 50 series, or data-centre cards (A100, H100, Blackwell), 16 to 24 GB+ for large humanoid populations |
| GPU driver | Linux 580.65.06 or later, Windows 580.88 or later | NVIDIA production branch driver matching the Isaac Sim release |
| Python | 3.10 (Isaac Sim 4.x) or 3.11 (Isaac Sim 5.x) | 3.11 |
Linux is the primary supported platform. Windows works for many workflows but lags Linux on some sensor and rendering features. Headless Linux servers with a virtual display are the standard configuration for cluster training.
The Isaac Lab 2.x install path is now pip install isaacsim followed by a git clone of the IsaacLab repository and a single ./isaaclab.sh --install command. Earlier Orbit-era installs required manually pointing the framework at an Isaac Sim install directory.
The pace of releases through 2024 and 2025 was unusually fast. A few items stand out.
The Mimic extension introduced in 2.1 lets practitioners record a small number of teleoperated demonstrations through Apple Vision Pro, a SpaceMouse, or a keyboard, then automatically expand those demonstrations into much larger imitation-learning datasets through trajectory perturbation and replay. The combination with Cosmos Transfer for visual augmentation produced the eye-catching 780,000-trajectory dataset behind the 2025 GR00T Blueprint.
FORGE and AutoMate in 2.2 added contact-rich industrial assembly tasks. These environments include peg-in-hole, gear meshing, and connector insertion, with realistic friction and tolerances that benchmark how well a policy can handle the kind of small, force-sensitive corrections that real assembly lines demand.
DexSuite in 2.3 expanded the dexterous manipulation benchmark with new tasks for the Allegro Hand, Shadow Hand, and Inspire hand on the Unitree G1. SkillGen added planning-based motion synthesis that hands off skill primitives to learned policies, mediated through cuRobo for fast collision-free motion planning.
Looking forward, the November 2025 paper points to a forthcoming Newton physics engine, a differentiable GPU physics back end being developed by NVIDIA, Google DeepMind, and Disney Research, that would let Isaac Lab do gradient-based learning through the physics simulation itself. An Isaac Lab 3.0 beta is in flight on Isaac Sim 6.0 with multi-backend physics, a pluggable renderer system, and Warp-native data pipelines.
Integration with NVIDIA Cosmos and the Isaac GR00T family of foundation models has tightened, to the point where the canonical NVIDIA story for humanoid robotics is now Cosmos plus Isaac Sim plus Isaac Lab plus GR00T as a single end-to-end stack from data generation to deployed policy.