> This page is for version v0.11.0 (latest) (default).
> For other versions, use one of these documentation indexes:
> - v0.11.0 (latest) (default): https://docs.nvidia.com/isaac-os/v0-11-0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/isaac-os/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/isaac-os/_mcp/server.

# Isaac OS

**Enable robotics developers to build and deploy enterprise-grade physical AI robot
applications with speed-of-light performance on NVIDIA Edge hardware.**

#### Speed of light

Optimized end to end for NVIDIA hardware, for deterministic real-time control next
to AI inference. [See the numbers](#performance).

#### Enterprise readiness

Built for production robots on the enterprise-supported
[Holoscan platform](https://holoscan.org).
[Ask about NVIDIA AI Enterprise](https://www.nvidia.com/en-us/data-center/products/ai-enterprise/contact-sales/).

#### ROS 2 interop

Connect your ROS 2 node with the [ROS 2 Bridge](/isaac-os/how-to/connect-a-ros-2-graph).
See how ROS 2 can map to Isaac OS [here](/isaac-os/explanation/isaac-os-and-ros-2).

## Full robotics stack

Isaac OS is the NVIDIA-native robotics stack for deploying humanoids, mobile robots
and manipulators. It brings together:

* **NVIDIA Edge hardware:** AGX and IGX [platforms](/isaac-os/reference/platforms).
* **NVIDIA system software:** JetPack, IGX OS, CUDA and TensorRT.
* **Holoscan-based middleware and reusable components:** compute graphs, message
  passing and GPU resource management for low-latency control alongside AI inference.
* **Developer resources:** reference applications, performance benchmarks, tutorials
  and agent skills.

[This repository](https://github.com/nvidia-isaac/isaac-os/tree/v0.11.0/) provides the reference applications and developer resources.

## Performance

**Real-time control and AI inference on the same chip.** Isaac OS partitions
CPU and GPU resources so control loops keep their deadlines and determinism while AI
policies and other heavy workloads share the hardware. Measured on Jetson AGX Thor with about
one hour of samples per condition:

#### 7× lower WBC tail latency, 13× less jitter

GR00T N1.7 runs at 5 Hz on the same GPU. With Green Context, WBC p99.9 latency:
**2.57 → 0.37 ms**, jitter **3.00 → 0.23 ms**.
[Full benchmark](/isaac-os/reference/applications/gr00t-n1-7-wbc-g1/benchmark)

#### 25× lower GPU control latency

1 kHz GPU IK, 1,024 candidates in parallel, under heavy GPU load. With Green Context
isolation: p99.99
**6.60 → 0.27 ms**, late commands **3.1 million → 0**.
[Full benchmark](/isaac-os/reference/applications/robot-control-benchmark)

#### Under 19 µs on the CPU

1 kHz CPU IK, one solve per cycle, on isolated PREEMPT_RT cores: p99.99
**under 19 µs** with or without stress-ng, **0** late commands.
[Full benchmark](/isaac-os/reference/applications/robot-control-benchmark)

## Try it now

**Run GR00T N1.7 and whole-body control on a simulated G1 in two commands.**
Start with [Getting started](/isaac-os/tutorials/getting-started) for host setup and
checkout. Run commands from the repository root; the simulation examples also use
[Isaac Sim setup](/isaac-os/how-to/set-up-isaac-sim). Start the two terminals in either order.

#### G1 + GR00T in Isaac Sim

*x86_64 only*

Run GR00T N1.7 with whole-body control to balance a G1 and move an apple onto a plate.
The first run builds five TensorRT engines and takes several minutes; later runs
reuse them. Set the same `DISPLAY` in both terminals to see the robot, or leave it
unset for a headless run.

**Terminal 1: simulator**

```bash
./isaac_sim run --scene operators/unitree_g1/assets/g1_static_apple_ucx.usda \
  -- --substeps 4
```

**Terminal 2: controller**

```bash
./isaac_os run gr00t_n1_7_wbc_g1 sim
```

[See details](/isaac-os/reference/applications/gr00t-n1-7-wbc-g1#isaac-sim).

#### SO-101 in Isaac Sim

*x86_64 only*

Move a simulated SO-101 arm with sinusoidal joint commands.

**Terminal 1: simulator**

```bash
./isaac_sim run --scene operators/so101/assets/so101_ucx_with_env.usda
```

**Terminal 2: controller**

```bash
./isaac_os run so101_sinus_motion sim
```

[See details](/isaac-os/reference/applications/so101-sinus-motion#isaac-sim).

#### SO-101 mock

Run the SO-101 control graph with synthetic inputs and inspect joint commands in
your terminal. No simulator or robot is needed; the run stops after 100 commands.

```bash
./isaac_os run so101_sinus_motion mock -- --message-count=100
```

[See details](/isaac-os/tutorials/getting-started).

---

## Go further

* **Evaluate:** explore [application results](/isaac-os/reference/applications/gr00t-n1-7-wbc-g1/benchmark#results)
  and [platform support](/isaac-os/reference/platforms).
* **Try:** follow [SO-101 reach deployment](/isaac-os/tutorials/so101-reach/deployment)
  or the [GR00T and whole-body control tutorial](/isaac-os/tutorials/gr00t-wbc).
* **Adapt:** [integrate components](/isaac-os/how-to/integrate-in-your-project),
  [add a robot embodiment](/isaac-os/how-to/add-a-robot-embodiment), or
  [exchange messages with ROS 2](/isaac-os/how-to/connect-a-ros-2-graph).