NVIDIA Isaac
GPU simulation (Isaac Sim), robot learning (Isaac Lab), accelerated ROS 2 packages (Isaac ROS) and Jetson onboard computers. Mostly free to download, but tied to NVIDIA hardware.
| Organisation | NVIDIA Corporation (Santa Clara, California) |
|---|---|
| Website | developer.nvidia.com/isaac |
| Licensing | Isaac Sim source under Apache-2.0 but needs the Omniverse Kit SDK and assets under NVIDIA terms; Isaac Lab under BSD-3-Clause; Isaac ROS under the NVIDIA Isaac ROS Software License, limited to systems with NVIDIA GPUs |
| Pricing | Isaac Sim, Isaac Lab and Isaac ROS are free downloads; Jetson AGX Thor developer kit launched at USD 3,499 in August 2025; GPUs at market price |
| Updated | 14 Nov 2026 |
Products and offerings #
| Name | Kind | Positioning |
|---|---|---|
| Isaac Sim | simulator | Omniverse-based robot simulator with PhysX, RTX sensor rendering, URDF and MJCF import and a ROS 2 bridge |
| Isaac Lab | framework | GPU-parallel reinforcement and imitation learning on top of Isaac Sim or the Newton backend |
| Newton | library | Warp-based physics backend with MuJoCo-Warp, VBD and MPM solvers |
| Isaac ROS | library | CUDA-accelerated ROS 2 packages for perception, localisation and manipulation |
| Jetson AGX Thor | controller | Onboard computer for robots, 128 GB memory, 2070 FP4 TFLOPS (sparse) according to NVIDIA |
Who they are
This hub is for teams that train policies in simulation, render synthetic sensor data, or run perception on the robot itself. NVIDIA Corporation was incorporated in April 1993 and is based in Santa Clara, California. It sells the GPUs that most robot learning runs on, and the Isaac brand is the robotics software it builds on top of them. NVIDIA was one of the founding Platinum members of the Open Source Robotics Alliance in March 2024, according to TechCrunch, so its ROS 2 work sits close to the ROS 2 hub. For background on the hardware itself, see the AI/ML Field Guide.
Tooling and APIs
Isaac Sim is the simulator. Version 6.2.0 was released on 14 October 2026. The source on GitHub is Apache-2.0, but building or running it pulls in the Omniverse Kit SDK, 3D models and textures under other NVIDIA terms. According to the system requirements page, the supported systems are Ubuntu 22.04 (deprecated) or 24.04, or Windows 11, and the minimum is a GeForce RTX 4080 or better with 16 GB of VRAM, and 32 GB of system memory. GPUs without RT cores, such as the A100 and H100, are not supported. On Arm, only DGX Spark is listed. There is no macOS build. Isaac Sim imports URDF, MJCF and CAD and stores everything as USD.
Isaac Lab is the learning framework, licensed BSD-3-Clause. Version 3.0.0 (20 October 2026) targets Isaac Sim 6.2, Python 3.12 and PyTorch 2.11 with CUDA 13.0, according to the release notes. The big change is that Isaac Sim is now optional. Tasks can run on Newton, a physics backend built on NVIDIA Warp that includes a MuJoCo-Warp solver, without the Omniverse application. The robot learning topic covers how these frameworks are used.
Isaac ROS is a set of ROS 2 packages that run perception and planning on the GPU. Version 5.0.0 (21 September 2026) moved to ROS 2 Lyrical Luth and ships Lyrical packages for Ubuntu 24.04 from NVIDIA’s own apt repository, according to the release notes. Note that this is not the Ubuntu 26.04 base that Open Robotics treats as Tier 1 for Lyrical.
Jetson modules are the onboard computers. The Jetson AGX Thor developer kit became generally available on 25 August 2025 starting at USD 3,499, according to NVIDIA’s announcement. The product page lists 128 GB of LPDDR5X and 2070 FP4 TFLOPS with sparsity for the T5000 module, measured at 130 W according to NVIDIA. The smaller T4000 has 64 GB and 1200 TFLOPS by the same measure.
Pain points: the driver, CUDA, PyTorch and Isaac versions must match, and the requirements page lists the exact driver builds NVIDIA tested. The minimum spec asks for 50 GB of SSD space. Isaac Sim needs a desktop-class RTX card, so a laptop or cloud A100 will not do.
Migration notes
- Quaternion order changed. Isaac Lab 2.x used (w, x, y, z). Isaac Lab 3.0 uses (x, y, z, w) in every API, configuration value and sensor output, and the identity quaternion is now (0, 0, 0, 1). That matches ROS messages and SciPy but no longer matches MuJoCo. Audit hard-coded rotations, reset states and observations when you port a 2.x task.
- Robot descriptions. URDF and MJCF go in through importers and come out as USD. Treat the USD as a generated file and keep the URDF or MJCF as the source of truth, or the two will drift. The URDF and MJCF cheat sheet covers the source formats.
- From MuJoCo. Newton’s MuJoCo-Warp solver gives Isaac Lab a MuJoCo-family contact model, but contact parameters, actuator models and timesteps are not guaranteed to carry over one to one. Re-tune gains and re-check sim-to-real transfer after any physics change.
- From Isaac Lab 2.x. Start from a fresh Python 3.12 environment. NVIDIA ships a migration guide with the release.
Watch for
- Hardware lock-in by licence. The Isaac ROS licence permits developing applications only for use in systems with NVIDIA GPUs. It forbids reverse engineering and forbids combining the software in a way that would bring it under an open-source licence. Check this before you mix Isaac ROS with GPL code.
- Export rules. The same licence states that the software is subject to US export law, including the Export Administration Regulations.
- Not for safety functions. The licence says the software is not tested or certified for systems whose failure could threaten human life. Keep safety in a certified controller.
- Fast release cadence. Isaac Lab and Isaac Sim each had several releases in 2026, and a major version broke the quaternion order. Pin versions per project and read release notes before upgrading.
- Maker figures. TFLOPS numbers are NVIDIA’s own peak figures under stated conditions. Measure your own model on the target module.