NVIDIA released Isaac Lab 3.0 as a general-availability release on October 19–20, 2026: tag v3.0.0 on GitHub and isaaclab 3.0.0 on PyPI. It targets Isaac Sim 6.2, Python 3.12 and PyTorch 2.11 with CUDA 13.0. The headline changes are selectable physics backends (PhysX or Newton, which includes MuJoCo-Warp), training without launching Isaac Sim for Newton workflows, one isaaclab train / play command for every RL library, and API changes such as quaternions switching to (x, y, z, w) order. Projects written for 2.x need a migration pass.
Isaac Lab 3.0 release timeline
| Release | Date (2026) | Isaac Sim | Status |
|---|---|---|---|
| v2.3.2 | February 2 (PyPI) | 5.1 (Python 3.11) | Previous stable line |
| v3.0.0-beta | March 17 | Not listed in the README table | Beta |
| v3.0.0-beta2 | June 17 | 6.0 | Beta |
| v3.0.0-EA | September 16 | 6.1 | Early access; GA targeted for late October |
| v3.0.0 | October 19 (PyPI), October 20 (GitHub) | 6.2 | General availability |
Dates come from the GitHub releases page and PyPI history; the Isaac Sim pairing comes from the Isaac Lab README and Isaac Sim’s 6.2 compatibility page, which lists support for Isaac Lab v3.0.0-GA. On October 24 the separate release-notes page on docs.nvidia.com still showed beta2.patch1 as the latest build, so read the GitHub release notes for current details.
What Isaac Lab 3.0 needs to run
- Python 3.12 only. The PyPI package requires Python 3.12 (3.13 is excluded).
- PyTorch 2.11 with CUDA 13.0, plus NVIDIA Warp 1.17, Newton 1.6.1 and RSL-RL 5.5.1 as pinned in the GA notes.
- Isaac Sim 6.2 when you use PhysX or RTX rendering through Omniverse Kit. Newton, OVPhysX and OVRTX workflows run “kit-less”, without starting Isaac Sim.
- Isaac Sim 6.2 hardware, per the requirements page: at least an RTX 4080 with 16 GB, 32 GB of RAM, Ubuntu 24.04 (22.04 still works but is deprecated) or Windows 11, and DGX OS 7 on DGX Spark. Data-center GPUs without RT cores, such as A100 and H100, are listed as unsupported for Isaac Sim. Our Isaac Sim 6 requirements and install guide covers drivers and install methods.
- Windows and DGX Spark. The GA notes list fixes for Windows wheel installs and for launch behavior on Linux aarch64 and DGX Spark. Isaac Lab itself publishes no separate OS support table.
What is new in Isaac Lab 3.0
- Separate physics, renderer and visualizer choices. Physics can be PhysX, Newton (with MuJoCo-Warp, VBD and MPM solvers) or OVPhysX, and the renderer is picked independently.
- Kit-less runs through Newton and the OV libraries, which lighten headless training machines. Our guide to choosing PhysX or Newton shows how to compare the two on one task.
- Warp-native data. Asset and sensor
.datafields return aProxyArray; call.torchor.warpto get the format you want. - One command line.
isaaclab trainandisaaclab playwork across RL libraries, with TorchRL integration and askrl-jaxextra added in GA. - Deformables: surface and volume deformables, and coupling between rigid, deformable and MPM objects.
- Data and deployment tools: Isaac Capture teleoperation, LEAPP policy export, and guidance on moving policies between PhysX and Newton.
- 17 browser-based interactive demos for trying environments before installing.
Breaking changes that affect 2.x code
| Area | Isaac Lab 2.x | Isaac Lab 3.0 |
|---|---|---|
| Quaternion order | (w, x, y, z) | (x, y, z, w) |
| Asset and sensor data | torch tensors | ProxyArray; use .torch or .warp |
| Launching scripts | AppLauncher | launch_simulation (AppLauncher kept with limited compatibility) |
| Config overrides | hydra_task_config, parse_overrides, apply_overrides | resolve_task_config |
| Actuator commands | On the articulation | Under robot.actuators; limit names changed |
| IMU sensor | Full-state IMU | Renamed PVA; a new IMU and JointWrenchSensor added |
| –headless flag | Turned off visualizers | No longer turns off configured visualizers; choose the visualizer explicitly |
| Video recording | gym.wrappers.RecordVideo | VideoRecorderCfg |
| Training entry points | train.py / play.py per RL library | isaaclab train / play |
| URDF and MJCF import | Previous importers | Rewritten; output layout changed |
| reset() | Positional arguments as before | env_ids is the first positional argument |
| Stable-Baselines3 actions | Previous default bounds | Default bounds [-1, 1] |
The full list, including deprecated aliases kept for 3.0, is in the v3.0.0 release notes. The EA notes also list the split of write_*_to_sim into _index and _mask variants.
How to move a 2.x project to Isaac Lab 3.0
- Create a fresh Python 3.12 environment with PyTorch 2.11 and CUDA 13.0. Install with
pip install isaaclabplus the extras you need (isaacsim,ov,ovphysx,ovrtx,teleop,mimic,torchrl,skrl-jaxorall), or clone the repository and useuv run; the oldisaaclab.shscript is deprecated. - Re-import robots with the new URDF and MJCF importers and check joint names, limits and the USD layout before training.
- Search your code for quaternion literals and conversions, since a silent (w, x, y, z) value now reads as a different rotation.
- Replace tensor access on
.datawith.torch, move actuator commands torobot.actuators, and switch scripts tolaunch_simulationandresolve_task_config. - Train one known task on PhysX with a fixed seed and compare reward curves with your 2.x baseline. Switch to Newton only after that matches.
NVIDIA’s migration guide has the step-by-step API mapping. If you are starting from zero, our Isaac Lab getting-started guide explains how Isaac Lab, Isaac Sim and the old Isaac Gym relate.
Stay on 2.3.2 or move to 3.0?
Isaac Lab 2.3.2 runs on Isaac Sim 5.1 and Python 3.11 and matches most published tutorials and course material, which helps a project near a deadline. New features such as Newton physics, kit-less runs and deformables arrive on the 3.0 line, so new projects should start there. The release notes give no end date for 2.x updates.
Isaac Lab 3.0 license
The Isaac Lab framework is BSD-3-Clause. The isaaclab_mimic extension is Apache 2.0 but depends on cuRobo, which has its own proprietary terms, and Isaac Sim is proprietary NVIDIA software (see the Isaac Lab repository). Newton is Apache-2.0 and a Linux Foundation project started by Disney Research, Google DeepMind and NVIDIA. For where Isaac Lab fits next to Cosmos, GR00T and Jetson Thor, see our NVIDIA physical AI stack overview.
Isaac Lab GitHub releases, PyPI and Isaac Sim 6.2 documentation checked October 24, 2026.