NVIDIA Isaac Lab

NVIDIA Isaac Lab is an open-source, GPU-accelerated framework for robot learning that trains robot control policies at scale by running thousands of physics simulations in parallel on a single graphics card, built on top of NVIDIA Isaac Sim.^[1][6] It is the official successor to NVIDIA Isaac Gym and a consolidation of the earlier Orbit framework, and it supports reinforcement learning, imitation learning, and motion planning, with photoreal RTX rendering and a manager-based modular API.^[1][7] The framework is licensed under the BSD-3-Clause license and hosted on GitHub at isaac-sim/IsaacLab.^[1][14] On 8 NVIDIA GPUs and 16,384 parallel environments, Isaac Lab reaches over 900,000 frames per second on the DextrAH dexterous-grasping teacher task and over 1.6 million frames per second on the Franka cabinet task, collapsing what would be weeks of CPU-based training into hours.^[5][8]

Isaac Lab 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.^[4][7] 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.^[5]

What is Isaac Lab used for?

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.^[6] 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.^[6]

The 2025 Isaac Lab paper describes the framework in one liftable sentence: "Isaac Lab combines high-fidelity GPU parallel physics, photorealistic rendering, and a modular, composable architecture for designing environments and training robot policies."^[5] Isaac Lab inherits the 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, and sensor models for cameras, LiDAR, IMUs and force-torque.^[6][2] 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, 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.^[5]

When was Isaac Lab released, and how does it relate to Isaac Gym?

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).^[5] NVIDIA describes Isaac Lab plainly as "the successor to Isaac Gym," subsuming all of Gym's features while adding Omniverse-based photoreal RTX visuals.^[7]

Scroll sideways for more →

How is Isaac Lab architected?

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.^[2] The physics engine is NVIDIA PhysX 5, configured to run thousands of independent worlds in parallel, each with its own contacts and articulations.^[2] The renderer is RTX, which supports both fast tiled rendering for camera observations and photoreal path tracing for synthetic data.^[2]

The code base is organised around a small number of long-lived concepts:

• Scene. A scene description in USD that contains terrain, lighting, robots, props, and sensors. Scenes can be authored visually in Isaac Sim or programmatically in Python.
• Articulation and rigid body APIs. Direct GPU-resident wrappers around PhysX articulations. Joint positions, velocities, root states, and contact forces are exposed as batched torch.Tensor objects, so the user never copies between CPU and GPU during training.
• Sensors. Camera (RGB, depth, semantic, instance), LiDAR via ray casting, IMU, contact sensor, force-torque sensor, frame transformer. Cameras can use either rasterised tiled rendering or path-traced RTX.
• Actuators. Models for ideal joint, DC motor, implicit position, and remote-actuated joint behaviour, with optional friction, torque limits, and effort and velocity caps.
• MDP terms. Modular reward, observation, termination, command, event, and curriculum modules. Each term is a small Python class that the manager calls every step. New tasks are usually built by composing existing terms rather than rewriting an environment from scratch.
• Direct and manager-based environments. A direct environment is a single Python class that owns its own step logic, similar to an old IsaacGymEnvs task. A manager-based environment delegates the MDP terms to managers, which is the recommended style for new work and which makes domain randomisation, multi-agent setups, and curriculum learning simpler.

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).^[2] 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).^[5]

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.^[2] 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.

Which reinforcement learning libraries does 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.^[2] The choice of library is left to the user.

Scroll sideways for more →

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.^[2]

How does Isaac Lab handle sim-to-real transfer?

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.^[8]

Domain randomisation is implemented through event terms in the manager-based environment.^[2] 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.^[2] 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.^[11]

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.^[2] 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.^[9] 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.^[9] NVIDIA reported that combining this synthetic data with real data improved GR00T N1's performance by 40 percent compared with using only real data.^[9]

Which robots does Isaac Lab support?

Isaac Lab ships with more than 16 pre-integrated robots and over 30 ready-to-train environments.^[2] The list keeps growing release by release; below is the picture as of the 2.3 release in late 2025.

Scroll sideways for more →

The framework is also a foundation for NVIDIA's own Isaac GR00T workflows, which target humanoid manipulation.^[9] 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.^[9]

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.^[11] Users can drop a new robot into Isaac Lab by writing a configuration class that describes its joints, actuators, and default initial state.^[2]

Notable use cases

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.^[6] 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.^[5]

Boston Dynamics Spot. Boston Dynamics has used Isaac Lab to train policies that are deployed directly for inference on Spot.^[8] The company has spoken publicly about how the GPU-parallel rollouts collapse weeks of CPU-based training into hours.^[8]

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.^[7] 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.^[9] At GTC, Huang demonstrated 1X's humanoid robot performing autonomous domestic tidying using a post-trained policy built on GR00T.^[7]

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.^[9]

University research. The framework underpins many recent results in dexterous manipulation, contact-rich assembly, and cross-embodiment locomotion. The 2025 Mittal et al. paper draws its contributors from NVIDIA, ETH Zurich, MIT, Stanford, CMU, Berkeley, and others.^[5]

How does Isaac Lab compare with other robot simulators?

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.

Scroll sideways for more →

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.

Software stack

The public surface of Isaac Lab is a Python API.^[2] Internally, the stack runs roughly as follows:

• Python user code calls into the isaaclab package, which exposes scene, asset, sensor, actuator, manager, and environment classes.
• The environment classes drive the omni.isaac.lab and broader Omniverse Kit extensions, which are loaded inside an Isaac Sim runtime.
• Isaac Sim hosts the USD stage, the PhysX 5 simulation, the RTX renderer, and the various sensor backends.
• All robot state is exposed back to Python as PyTorch tensors that share GPU memory with the simulation. Policy networks, written in PyTorch and accelerated with CUDA and cuDNN, read these tensors directly without round-tripping through the CPU.
• For multi-GPU or multi-node training, PyTorch's 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.^[2] Asset libraries (USD scenes, robot meshes, materials) are downloaded from NVIDIA's Nucleus servers and cached locally.^[2]

Is Isaac Lab open source, and what hardware does it need?

Isaac Lab itself is open source under the BSD-3-Clause license.^[14] The isaaclab_mimic extension is Apache 2.0. Both can be cloned, modified, and redistributed with attribution.^[14]

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.^[12] Internal R&D use is free of charge.^[12] Redistributing Isaac Sim with Omniverse Kit, or delivering it as a service to third parties, requires an NVIDIA AI Enterprise licence.^[12] 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.^[12]

Hardware and operating system requirements as of Isaac Lab 2.3 and Isaac Sim 5.1:^[2]

Scroll sideways for more →

Linux is the primary supported platform.^[2] Windows works for many workflows but lags Linux on some sensor and rendering features.^[2] 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.^[2] Earlier Orbit-era installs required manually pointing the framework at an Isaac Sim install directory.^[13]

Recent developments

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.^[11] The combination with Cosmos Transfer for visual augmentation produced the eye-catching 780,000-trajectory dataset behind the 2025 GR00T Blueprint.^[9]

FORGE and AutoMate in 2.2 added contact-rich industrial assembly tasks.^[10] 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.^[10]

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.^[11] SkillGen added planning-based motion synthesis that hands off skill primitives to learned policies, mediated through cuRobo for fast collision-free motion planning.^[11]

Newton physics engine. The November 2025 paper points to Newton, a differentiable, GPU-accelerated physics engine built on NVIDIA Warp and OpenUSD that lets Isaac Lab do gradient-based learning through the physics simulation itself.^[5] Newton is jointly developed by NVIDIA, Google DeepMind, and Disney Research, was first unveiled at GTC in March 2025, and was contributed to the Linux Foundation in September 2025 as an open-source project.^[15] It integrates MuJoCo Warp and reports up to 70x acceleration for humanoid simulation and 100x for in-hand manipulation; Disney plans to use Newton for its next-generation BDX robotic characters.^[15] 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.^[11]

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.^[9]

References

1. Isaac Lab GitHub repository: https://github.com/isaac-sim/IsaacLab ↩
2. Isaac Lab documentation: https://isaac-sim.github.io/IsaacLab/ ↩
3. NVIDIA Developer page for Isaac Lab: https://developer.nvidia.com/isaac/lab
4. Mittal, M. et al. (2023). Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments. IEEE Robotics and Automation Letters, vol. 8, no. 6, pp. 3740 to 3747. arXiv:2301.04195. DOI 10.1109/LRA.2023.3270034. ↩
5. Mittal, M. et al. (2025). Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. arXiv:2511.04831. https://arxiv.org/abs/2511.04831 ↩
6. Makoviychuk, V. et al. (2021). Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning. NeurIPS Datasets and Benchmarks Track. arXiv:2108.10470. ↩
7. NVIDIA Newsroom (18 March 2024). NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots and Major Isaac Robotics Platform Update. https://nvidianews.nvidia.com/news/foundation-model-isaac-robotics-platform ↩
8. NVIDIA Technical Blog. Fast-Track Robot Learning in Simulation Using NVIDIA Isaac Lab. https://developer.nvidia.com/blog/fast-track-robot-learning-in-simulation-using-nvidia-isaac-lab/ ↩
9. NVIDIA Technical Blog. Building a Synthetic Motion Generation Pipeline for Humanoid Robot Learning. https://developer.nvidia.com/blog/building-a-synthetic-motion-generation-pipeline-for-humanoid-robot-learning/ ↩
10. NVIDIA Technical Blog. Announcing General Availability for NVIDIA Isaac Sim 5.0 and NVIDIA Isaac Lab 2.2. https://developer.nvidia.com/blog/isaac-sim-and-isaac-lab-are-now-available-for-early-developer-preview/ ↩
11. Isaac Lab Release Notes: https://isaac-sim.github.io/IsaacLab/main/source/refs/release_notes.html ↩
12. Isaac Sim License FAQ: https://docs.isaacsim.omniverse.nvidia.com/6.0.0/common/license-faq.html ↩
13. ORBIT project page: https://isaac-orbit.github.io/ ↩
14. Isaac Lab BSD-3-Clause license: https://github.com/isaac-sim/IsaacLab/blob/main/LICENSE ↩
15. Linux Foundation (September 2025). Linux Foundation Announces Contribution of Newton by Disney Research, Google DeepMind and NVIDIA to Accelerate Open Robot Learning. https://www.linuxfoundation.org/press/linux-foundation-announces-contribution-of-newton-by-disney-research-google-deepmind-and-nvidia-to-accelerate-open-robot-learning ↩