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 GPUs.^[1][6] Originally built on NVIDIA Isaac Sim, it can also run without Isaac Sim through the kit-less backends available in Isaac Lab 3.0.^[27] Its tagged 2.x releases ran only on NVIDIA PhysX 5, although NVIDIA published an experimental Newton backend for Isaac Lab as a beta on 29 September 2025;^[26] the Isaac Lab 3.0 beta published on 17 March 2026 made that a formal multi-backend physics architecture with PhysX as the beta's default and Newton, the open-source engine co-developed by NVIDIA, Google DeepMind, and Disney Research, as a selectable alternative that can also run without Isaac Sim.^[20][21] 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]

The official v3.0.0 release was published on GitHub on 20 October 2026. It is the generally available 3.0 release, following the March beta and September Early Access release.^[27][28] NVIDIA Robotics announced it on X on 22 October 2026 as bringing a "new modular and lightweight architecture", naming Newton integration, "the new OVRTX renderer", RL post-training and LEAPP deployment as its headline additions.^[35] Its backend choices, task-specific compatibility and migration requirements are described below.

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] In the original PhysX backend, PhysX 5 simulates rigid bodies, articulated systems, and deformable objects in batched form, and observations are read directly from PhysX buffers as PyTorch tensors. Since the 3.0 beta the same loop can run on the Newton backend, where the MuJoCo Warp solver steps the scene and the beta exposed asset data as Warp arrays that convert to PyTorch with wp.to_torch().^[20] The final 3.0 release uses ProxyArray wrappers, with explicit .torch and .warp accessors.^[30] 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]

How is Isaac Lab architected?

Isaac Lab 2.x 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] In 3.0, Isaac Sim is optional, and the physics, renderer and visualizer are separate configuration choices.^[27][29] Through the tagged Isaac Lab 2.x releases the physics engine was NVIDIA PhysX 5 (an experimental Newton beta backend was published alongside them in September 2025^[26]), configured to run thousands of independent worlds in parallel, each with its own contacts and articulations, and PhysX was also the default in the first 3.0 beta.^[2][20] In the final 3.0 release, each task declares its supported presets and default; there is no need to assume that every environment uses the same backend.^[29] The renderer is RTX, which supports both fast tiled rendering for camera observations and photoreal path tracing for synthetic data.^[2]

The Isaac Lab 3.0 beta (v3.0.0-beta, published on GitHub on 17 March 2026 and built on Isaac Sim 6.0) replaced the single-engine design with what its release notes call a "factory-based multi-backend architecture". Asset and sensor classes such as Articulation, RigidObject, and ContactSensor are now abstract base classes with backend-specific implementations in separate packages: isaaclab_physx, the full PhysX backend and the default, covering deformable objects, surface grippers, contact sensors, IMUs, and frame transformers; and isaaclab_newton, a Newton backend built on MuJoCo Warp that supports the MJWarp, XPBD, and Featherstone solvers with CUDA-graph acceleration. Existing isaaclab.assets and isaaclab.sensors imports keep working because the factory dispatches to the active backend at runtime.^[20] A backend is chosen per run through presets (presets=newton on the training command line; omitting it keeps PhysX), and in a Newton run a NewtonCfg takes the place of the PhysxCfg for every simulation parameter except the time step, wrapping a solver configuration such as MJWarpSolverCfg.^[20][23] The Newton backend can run in a kit-less mode with no Isaac Sim runtime at all, paired with the Newton Warp renderer or the OVRTX pipeline instead of Isaac RTX.^[20] NVIDIA's Newton 1.0 blog of 16 March 2026 described the Isaac Sim 6.0 and Isaac Lab 3.0 integrations as "early access releases" at that point, Isaac Sim 6.0's own release notes list Newton as "experimental support", and the beta-era Isaac Lab documentation filed the Newton integration under "Experimental Features", focused primarily on the MuJoCo Warp solver; that overview listed a limited set of classic RL and flat-terrain locomotion examples, although its training-environments list and the beta release notes also covered manipulation presets.^[22][24][25] The documentation says the refactor was written so that "any other physics engine" can be plugged in, that many backend-neutral imports would continue working, and that policies have been transferred from Newton to PhysX and back, with a Newton-trained locomotion policy deployed on a G1 humanoid robot.^[21][22] NVIDIA's product page now pitches the framework as training "with higher-fidelity physics using Newton, PhysX, or any physics engine".^[3] The beta also switched quaternions from wxyz to xyzw to match PhysX, Warp, and Newton conventions, and moved the .data.* properties of assets and sensors from PyTorch tensors to NVIDIA Warp arrays.^[20][21]

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 the active backend's articulations (PhysX by default, Newton when selected in 3.0). Joint positions, velocities, root states, and contact forces are exposed as batched tensors, torch.Tensor objects through 2.x and ProxyArray wrappers in the final 3.0 release, with zero-copy Torch/Warp access.^[20][30]
• 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.

The table reflects the 2.x line. The 3.0 release notes list six integrated RL libraries behind the unified isaaclab train command: RSL-RL, RL-Games, skrl (with PyTorch or JAX), Stable-Baselines3, RLinf and TorchRL, with TorchRL integration and PPO examples added between Early Access and general availability.^[27]

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.

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]

Microduck. An NVIDIA Technical Blog tutorial of 23 October 2026 adapted a pretrained locomotion policy from the Isaac Lab Microduck example environment to ramps and obstacles, using head-camera depth and robot-state observations and 4,096 parallel environments on an NVIDIA RTX PRO 6000 Blackwell Workstation Edition GPU, with Newton simulating the robot's physics. The post describes using Isaac Lab's Better Actuator Models (BAM) to model servo command delays, backlash and voltage drops before sim-to-real transfer, and shows a policy trained in Isaac Lab walking on the physical robot.^[41]

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.

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] For 2.x and Isaac Sim-based workflows, the stack runs roughly as follows. Version 3.0 also supports the standalone paths described in the following section.

• 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.
• In the default configuration Isaac Sim hosts the USD stage, the PhysX 5 simulation, the RTX renderer, and the various sensor backends; in the 3.0 beta's kit-less mode the Newton backend steps physics and the Newton Warp or OVRTX renderer draws camera observations without an Isaac Sim runtime.^[20]
• In 2.x, 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, required for Isaac Lab 2.x but optional in 3.0, has separate licensing terms.^[27] 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]

For this Isaac Sim 5.1-based configuration, GPUs without RT cores, including A100 and H100, are explicitly unsupported; the requirements do not describe every kit-less 3.0 configuration.^[34]

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 3.0 beta moves the targets to Ubuntu 22.04 or 24.04, Python 3.12, and PyTorch 2.10.0, with Isaac Sim 6.0 optional when the Newton backend and a non-Kit visualizer are used.^[16][22]

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]

Isaac Lab 3.0 general availability

Version 3.0 can change a supported task's physics, renderer or visualizer without rewriting task logic; combinations and defaults are task-specific.^[29] The release targets Python 3.12, PyTorch 2.11 with CUDA 13.0, Warp 1.17, Newton 1.6.1 and RSL-RL 5.5.1. Isaac Sim-based workflows target Isaac Sim 6.2.^[27]

Supported combinations are listed by uv run isaaclab list_envs --show_presets; task --help lists choices. OVPhysX and OVRTX cannot share a process with Isaac Sim/Kit.^[29]

Newton solvers and kit-less execution

According to the release notes, most Isaac Lab 3.0 workflows can run without installing or launching Isaac Sim, using Newton physics with the MuJoCo-Warp, VBD and MPM solvers, coupled solvers, and an experimental Kamino solver.^[27] NVIDIA's announcement thread assigns each solver a role: MJWarp for articulated rigid bodies, MPM for particles, VBD for deformable and multiphysics simulation, and Kamino for closed-chain linkages, with solvers combinable through coupling.^[35] The "OV libraries", OVPhysX for PhysX and OVRTX for RTX rendering, are standalone libraries installed as optional extras, and the full Isaac Sim stack is needed only for features such as Isaac Sim PhysX or RTX, the ROS bridge, XR teleoperation, livestreaming or Kit tooling.^[27] The thread also says NVIDIA's SONIC whole-body controller runs across different physics engines and transfers to real hardware.^[35]

Rendering and visualization

The 3.0 documentation lists three camera-rendering backends.^[37]

ovrtx itself is a separately distributed NVIDIA C and Python library for embedding Omniverse RTX sensor simulation (camera, lidar, radar and other sensors) into applications. Its package page marks it as a pre-release whose API may change and places it under the NVIDIA Software License Agreement rather than Isaac Lab's BSD-3-Clause license.^[36] Visualizers are selected separately from renderers: the release notes list newton_gl, an experimental newton_rtx, the browser-based Viser and Rerun viewers, and kit.^[27] In its announcement thread NVIDIA also said Isaac Lab supports 3D Gaussian rendering for bringing reconstructed real-world environments into simulation, that "skinning gaussians" for deformable simulation is in an early access branch, and that NVIDIA Cosmos 3-Nano Transfer can turn simulated depth into photorealistic video guided by language prompts.^[35]

Policy export with LEAPP

LEAPP stands for Lightweight Export Annotations for Policy Pipelines.^[38][39] In Isaac Lab it exports a trained policy together with its observation processing, action processing and recurrent-state handling, so that deployment does not have to reimplement those steps by hand. The export directory contains model files (.onnx by default, or .pt), metadata including the policy execution frequency, a .safetensors file of initial feedback state such as recurrent hidden state or the last action, and an optional .png graph of the pipeline.^[38] Exported policies can be run in simulation with LeappDeploymentEnv, on a robot with Isaac ROS Deploy, or in a custom runtime that reads the exported YAML. Manager-based policies trained with RSL-RL, RL-Games, skrl or Stable-Baselines3 are supported, while Direct-workflow export covers RSL-RL policies after LEAPP annotations are added.^[38] The GA release added camera inputs and frame history for LEAPP.^[27] LEAPP is also a standalone NVIDIA library ("Isaac LEAPP") that exports annotated pipelines to ONNX, TorchScript or PyTorch ExportedProgram models; its PyPI package is Apache-2.0 licensed.^[39]

Interactive browser demos and teleoperation

The 3.0 documentation includes 17 interactive browser demos that run a Newton simulation locally on the user's CPU through WebAssembly: solver-tuning scenes (VBD stiffness, cloth and cable bending, MJWarp friction, joint PD tuning, MPM materials and a domino cascade), pretrained policies (Cartpole, G1 velocity control with the WBC-AGILE policy, ANYmal-D, MicroDuck, and several Franka reach, lift, drawer, soft-body and cloth tasks), and a scripted Franka cube-stacking scene. The documentation cautions that the simplified scenes can behave differently from native tasks.^[40] For teleoperation, 3.0 integrates Isaac Capture through isaaclab_teleop, with pipeline-based retargeting, CloudXR and MCAP record and replay, and deprecates the native isaaclab.devices.openxr stack.^[27]

Stabilization work between Early Access and GA covered Windows wheel installation, Linux aarch64 and DGX Spark launch behavior, and RTX multi-GPU training with PyTorch 2.11 and CUDA 13.^[27]

Migration from 2.x

The 3.0 migration guide describes changes to data access, rotations and command-line entry points. It is a major-version migration rather than a drop-in replacement for every 2.x script.^[30]

Coding-agent support and VLA post-training

Isaac Lab's agent skills are repository-maintained Markdown workflow instructions for installation, environment creation, backend selection, training, debugging and migration. Separate developer skills cover contribution conventions and review tasks. The documented discovery aliases let OpenAI Codex and Claude Code find the same maintained skills.^[31] These are instructions for coding tools working on Isaac Lab, not claims that a trained robot operates through a coding assistant.

The documentation also provides an experimental RLinf workflow for reinforcement-learning post-training of vision-language-action models. NVIDIA's release thread refers to this capability as "Isaac RL Post Training (RLPT)", which it says lets large models such as GR00T refine pretrained skills through parallel simulation.^[35] It includes GR00T N1.5 and N1.7 setup instructions and supports distributed training with FSDP. The documented integration is Linux-only, and the two GR00T generations require separate environments because their gr00t packages have incompatible APIs.^[32]

October 2026 Isaac platform announcement

NVIDIA's 21 October 2026 GTC Berlin announcement placed Isaac Lab 3.0 within the broader Isaac platform. It described open-source DGX Station reference workflows for GR00T reinforcement learning, dexterous manipulation in Isaac Lab and Cosmos post-training. The same announcement introduced Isaac OS as a runtime framework and said the Agile ONE SN reference humanoid was scheduled for early 2027. Those are separate platform components and availability statements, not additional Isaac Lab releases.^[33] The announcement described Isaac Lab 3.0 as an "open source, agent-ready learning framework" and named several users: Hugging Face and Pollen Robotics are using Isaac Lab, Isaac Sim and Newton to build Microduck; Flexion is using Isaac Lab and Newton for its Reflect model; NEURA Robotics is using Isaac Sim and Isaac Lab for its Neuron AI foundation model; and Schaeffler is contributing actuator technology as an "exclusive foundational component" for Newton.^[33]

Recent developments

The pace of releases through 2024 and 2025 was unusually fast. A few items stand out.

The Mimic extension, introduced in 2.0 and extended in 2.1 with Apple Vision Pro teleoperation and bi-manual workflows,^[42] 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 on Isaac Sim 6.0 with multi-backend physics, a pluggable renderer system, and Warp-native data pipelines was published on GitHub as v3.0.0-beta on 17 March 2026, followed by v3.0.0-beta2 on 17 June 2026 and a v3.0.0-beta2.patch1 on 2 July 2026 for Isaac Sim 6.0.1.^[11][16] The Beta 2 notes call it a stabilization release with "additional features and improvements on Newton support (VBD, solver coupling, Kamino, rough terrain, sensors)" alongside multi-backend physics, simplified training and installation commands, and kit-less workflows.^[16]

Newton Learning Path course (September 2026). On 4 September 2026, NVIDIA Robotics announced a free, self-paced course on NVIDIA's Physical AI Learning site, "Accelerating Robot Learning With Isaac Lab and Newton," whose second module (about 90 minutes) runs Newton as the physics engine of an Isaac Lab task and was tested against Isaac Lab v3.0.0-beta2 on Linux x86-64 with a supported NVIDIA GPU.^[17][18] In that module Newton is configured through a three-layer stack, an MJWarpSolverCfg for the MuJoCo Warp solver wrapped by a NewtonCfg wrapped by SimulationCfg, and a Direct RL environment (Template-Franka-Cube-Direct-v0) teaches a Franka arm to lift a cube with a PPO policy trained through RSL-RL; the course ships a trained checkpoint so the policy can be played back without training.^[18][19] The course can run locally or in NVIDIA Brev cloud environments, which the setup page says are billed while running, including when idle.^[18]

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. ^1 ^2 ^3Isaac Lab GitHub repository: github.com/...IsaacLab
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20Isaac Lab documentation: isaac-sim.github.io/IsaacLab
3. ^NVIDIA Developer page for Isaac Lab: developer.nvidia.com/...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. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Mittal, M. et al. (2025). Isaac Lab: A GPU-Accelerated Simulation Framework for Multi-Modal Robot Learning. arXiv:2511.04831. arxiv.org/...2511.04831
6. ^1 ^2 ^3 ^4 ^5Makoviychuk, V. et al. (2021). Isaac Gym: High Performance GPU-Based Physics Simulation For Robot Learning. NeurIPS Datasets and Benchmarks Track. arXiv:2108.10470.
7. ^1 ^2 ^3 ^4 ^5NVIDIA Newsroom (18 March 2024). NVIDIA Announces Project GR00T Foundation Model for Humanoid Robots and Major Isaac Robotics Platform Update. nvidianews.nvidia.com/...l-isaac-robotics-platform
8. ^1 ^2 ^3 ^4NVIDIA Technical Blog. Fast-Track Robot Learning in Simulation Using NVIDIA Isaac Lab. developer.nvidia.com/...ion-using-nvidia-isaac-lab
9. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9NVIDIA Technical Blog. Building a Synthetic Motion Generation Pipeline for Humanoid Robot Learning. developer.nvidia.com/...or-humanoid-robot-learning
10. ^1 ^2NVIDIA Technical Blog. Announcing General Availability for NVIDIA Isaac Sim 5.0 and NVIDIA Isaac Lab 2.2. developer.nvidia.com/...or-early-developer-preview
11. ^1 ^2 ^3 ^4 ^5 ^6Isaac Lab Release Notes: isaac-sim.github.io/...release_notes
12. ^1 ^2 ^3 ^4Isaac Sim License FAQ: docs.isaacsim.omniverse.nvidia.com/...license-faq
13. ^ORBIT project page: isaac-orbit.github.io
14. ^1 ^2 ^3Isaac Lab BSD-3-Clause license: github.com/...LICENSE
15. ^1 ^2Linux Foundation (September 2025). Linux Foundation Announces Contribution of Newton by Disney Research, Google DeepMind and NVIDIA to Accelerate Open Robot Learning. linuxfoundation.org/...elerate-open-robot-learning
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Isaac Lab GitHub release v3.0.0-beta2 (17 June 2026); the same releases page lists v3.0.0-beta (17 March 2026) and v3.0.0-beta2.patch1 (2 July 2026). github.com/...v3.0.0-beta2
17. ^NVIDIA Robotics (@NVIDIARobotics), X post announcing the Newton Learning Path, 4 September 2026. x.com/...2095950021636046876
18. ^1 ^2 ^3NVIDIA Physical AI Learning, "Accelerating Robot Learning With Isaac Lab and Newton": Getting Started page (tested versions, NVIDIA Brev and local setup) and the Isaac Lab and Newton module overview. docs.nvidia.com/...setup and docs.nvidia.com/...overview
19. ^NVIDIA Physical AI Learning, "Scene and Simulation Setup" and "Training and Playback" lessons of the Isaac Lab and Newton module. docs.nvidia.com/...scene-and-simulation and docs.nvidia.com/...training-and-playback
20. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10isaac-sim/IsaacLab, "Isaac Lab 3.0 Beta" (v3.0.0-beta), GitHub release, 17 March 2026. github.com/...v3.0.0-beta
21. ^1 ^2 ^3Isaac Lab documentation, "Isaac Lab 3.0 Beta" (Experimental Features: Newton Physics Integration). isaac-sim.github.io/...isaaclab_newton-beta-2
22. ^1 ^2 ^3Isaac Lab documentation, "Newton Physics Integration" overview, "Installation," and "Training Environments" pages. isaac-sim.github.io/...newton-physics-integration, isaac-sim.github.io/...installation, and isaac-sim.github.io/...training-environments
23. ^Isaac Lab documentation, "Solver Transitioning" (NewtonCfg replacing PhysxCfg; MJWarpSolverCfg). isaac-sim.github.io/...solver-transitioning
24. ^1 ^2NVIDIA Technical Blog, "Newton Adds Contact-Rich Manipulation and Locomotion Capabilities for Industrial Robotics," Philipp Reist, Miguel Zamora Mora, JC Chang, Rishabh Chadha and Mohammad Mohajerani, 16 March 2026. developer.nvidia.com/...es-for-industrial-robotics
25. ^NVIDIA Isaac Sim documentation, "Release Notes," 6.0.0 GA ("Added experimental support for the Newton physics engine"). docs.isaacsim.omniverse.nvidia.com/...release_notes
26. ^1 ^2 ^3Mohammad Mohajerani, Aravindh Shanmuganathan, Eric Heiden, Anka Chen, Gilles Daviet, Milad Rakhsha and Antoine Richard, "Train a Quadruped Locomotion Policy and Simulate Cloth Manipulation with NVIDIA Isaac Lab and Newton," NVIDIA Technical Blog, September 29, 2025. developer.nvidia.com/...vidia-isaac-lab-and-newton
27. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14isaac-sim/IsaacLab, "Isaac Lab 3.0," GitHub release v3.0.0, 20 October 2026. github.com/...v3.0.0
28. ^1 ^2isaac-sim/IsaacLab, "Isaac Lab 3.0 Early Access," GitHub release v3.0.0-EA, 16 September 2026. github.com/...v3.0.0-EA
29. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Isaac Lab documentation, "Backends and Presets," accessed 23 October 2026. isaac-sim.github.io/...backends_and_presets
30. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Isaac Lab documentation, "Migrating To 3.0," accessed 23 October 2026. isaac-sim.github.io/...migrating
31. ^Isaac Lab documentation, "Agent Skills," accessed 23 October 2026. isaac-sim.github.io/...agent_skills
32. ^Isaac Lab documentation, "RL Post-Training for VLA Models," experimental features, accessed 23 October 2026. isaac-sim.github.io/...rlinf_vla_posttraining
33. ^1 ^2NVIDIA Newsroom, "NVIDIA Advances Isaac Platform for General-Purpose Robotics With New Reference Humanoid, Frameworks, Simulation Pipelines and Data," 21 October 2026. nvidianews.nvidia.com/...lation-pipelines-and-data
34. ^1 ^2NVIDIA Isaac Sim documentation, "Isaac Sim Requirements," version 5.1.0. docs.isaacsim.omniverse.nvidia.com/...requirements
35. ^1 ^2 ^3 ^4 ^5NVIDIA Robotics (@NVIDIARobotics), seven-post X thread announcing the official release of Isaac Lab 3.0, 22 October 2026 (UTC). x.com/...2113407332747293131 (posts 2/7 to 7/7: x.com/...2113407335133790314, x.com/...2113407337314795961, x.com/...2113407339596566935, x.com/...2113407341668524364, x.com/...2113407343891468290, x.com/...2113407346085089675)
36. ^NVIDIA, ovrtx package page ("Omniverse RTX"), PyPI, accessed 25 October 2026. pypi.org/...ovrtx ; repository github.com/...ovrtx
37. ^1 ^2 ^3 ^4Isaac Lab documentation (v3.0.0), "Renderers," accessed 25 October 2026. isaac-sim.github.io/...renderers
38. ^1 ^2 ^3Isaac Lab documentation (v3.0.0), "Exporting Policies with LEAPP" and "Deploy Exported Policies with LEAPP," accessed 25 October 2026. isaac-sim.github.io/...exporting_policies_with_leapp and isaac-sim.github.io/...xported_policies_with_leapp
39. ^1 ^2NVIDIA, "Isaac LEAPP: machine learning pipeline export" documentation, and leapp package page on PyPI, accessed 25 October 2026. nvidia-isaac.github.io/leapp and pypi.org/...leapp
40. ^Isaac Lab documentation (v3.0.0), "Interactive Demos," accessed 25 October 2026. isaac-sim.github.io/...interactive_demos
41. ^Gabriele Leone, Seun Doherty, Asier Arranz, Johnny Núñez and Asawaree Bhide, "Using AI Agents to Teach Robots Like Microduck to Walk in Simulation," NVIDIA Technical Blog, 23 October 2026. developer.nvidia.com/...duck-to-walk-in-simulation
42. ^1 ^2isaac-sim/IsaacLab, GitHub releases v2.0.0 (30 January 2025) and v2.1.0 (24 April 2025). github.com/...v2.0.0 and github.com/...v2.1.0