NVIDIA Isaac Sim
GPU-accelerated robotics simulator for physics-accurate testing, synthetic data generation and large-scale policy training, with ROS 2 bridging and sim-to-real workflows.
Overview
Isaac Sim is NVIDIA's robotics simulator, built on the Omniverse Kit and used for physics-accurate testing, synthetic data generation and large-scale policy training. Version 6.0 went generally available on 4 June 2026 after an early developer build at GTC in March. The biggest structural change is a second physics backend: Newton, a GPU-native engine, now runs alongside PhysX, and the URDF and MJCF importers emit assets you can switch between the two. The old PyTorch-based Core API is deprecated in favour of a Warp-based Core Experimental API built to work across both.
Most of the real work happens through Isaac Lab, the BSD-3 licensed robot learning framework layered on top. It exposes a Gym-style interface into PyTorch or JAX and runs thousands of parallel environments on one GPU, which is how modern locomotion and manipulation policies get trained in hours rather than months. There is a ROS 2 bridge for software-in-the-loop testing, NuRec Gaussian splatting for rebuilding real scenes as assets, and the whole stack feeds NVIDIA's Isaac GR00T humanoid work.
Cost is zero for internal development, which is unusual for something this capable, but the hardware is not. Plan on an RTX card with 16 GB of VRAM and 32 GB of system memory as a floor, and expect complex scenes to want more. Anyone redistributing Isaac Sim inside a product or offering it as a service needs an NVIDIA AI Enterprise licence for the Kit components, reported by developers at roughly $4,500 per GPU per year. Strong fit for a robotics team with real hardware and a sim-to-real problem. Poor fit if you want a lightweight kinematics sandbox, where Gazebo or MuJoCo cost far less time to stand up.
Key Features
- ✓ Version 6.0 reached general availability on 4 June 2026 and ships two physics backends, PhysX and the GPU-native Newton engine, switchable per asset
- ✓ Isaac Lab, the BSD-3 robot learning framework on top, runs thousands of parallel environments on one GPU behind a Gym-style interface into PyTorch or JAX
- ✓ URDF and MJCF importers plus setup tooling: Gains Tuner, Robot Assembler, Robot Inspector, Robot Poser and a self-collision detector, all Newton-aware
- ✓ Synthetic data generation with domain randomisation, and NuRec Gaussian splatting to rebuild real captured scenes as simulation assets
- ✓ ROS 2 bridge for software-in-the-loop testing against the same stack that runs on the physical robot, feeding the Isaac GR00T policy pipeline
Where it holds
- • Nothing else touches its parallel-environment throughput for reinforcement learning on a single GPU
- • Apache 2.0 source on GitHub and zero cost for internal development, unusual for software this capable
- • Importers and setup tools get you from a URDF to a trainable scene in about a day instead of a fortnight
- • The Newton backend is a real escape hatch when PhysX contact behaviour does not match how your hardware actually behaves
Where it breaks
- • The hardware floor is unforgiving. An RTX card with 16 GB VRAM and 32 GB system RAM is the practical minimum, and 6 GB cards will not launch it at all
- • 6.0 deprecated the PyTorch-based Core API for a Warp-based Core Experimental API, so anything written against 4.x needs porting
- • Shipping it inside a product or offering it as a service means an NVIDIA AI Enterprise licence for the Kit components, reported around $4,500 per GPU per year
- • The sim-to-real gap is still your problem. Contact-rich manipulation transfers far worse than locomotion, and no amount of randomisation fully closes it
My Take
This is for teams that already own a robot and a GPU budget, not for someone curious about robot learning on a laptop. The 6.0 general availability build landed 4 June 2026 and brought Newton in as a second physics backend alongside PhysX, a meaningful hedge when contact dynamics do not match your rig. The cost is churn: the PyTorch Core API is deprecated in favour of a Warp-based one, so 4.x code needs porting, and the 16 GB VRAM floor quietly rules out a lot of existing workstations. Gazebo remains the cheaper answer for plain kinematic checks.
Quick Info
- Pricing:
- free
- Openness:
- Proprietary
- Starting at:
- Free to download and use for internal R&D with an NVIDIA Developer Program membership. The Isaac Sim source is Apache 2.0 on GitHub, but the bundled Omniverse Kit SDK, 3D models and textures sit under a separate proprietary NVIDIA licence. Redistributing Isaac Sim inside a product, or delivering it as a service to third parties, requires an NVIDIA AI Enterprise licence for the Kit components, which developers on NVIDIA's own forums put at roughly $4,500 per GPU per year. Selling only simulation outputs such as data or reports needs no licence. Isaac Lab is BSD-3.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where NVIDIA Isaac Sim holds and where it breaks. The fastest check is your own workload.
Visit website ↗Alternatives to NVIDIA Isaac Sim
Genesis World
open-sourceOpen-source multi-physics simulator for robot learning. Rigid, FEM, MPM, SPH and PBD solvers share one Python scene, with a photoreal renderer and heavy GPU batching.
MuJoCo
open-sourceGoogle DeepMind's rigid-body physics engine for robotics. Contact-rich dynamics, MJCF models, plus MJX and MuJoCo Warp for batched GPU rollouts during policy training.
BlenderProc
open-sourceGPL-3.0 Blender pipeline from DLR that renders photorealistic synthetic training images with RGB, depth, normals, segmentation and exact 6D pose, exporting to BOP and COCO.