> ## Documentation Index
> Fetch the complete documentation index at: https://dragonwingdocs.qualcomm.com/llms.txt
> Use this file to discover all available pages before exploring further.
# QRB ROS Ecosystem
**QRB ROS** is Qualcomm's upstream collection of ROS 2 Jazzy packages for robotics platforms — currently the Qualcomm Dragonwing™ IQ‑9075 EVK and IQ‑8 (Beta) EVK. This page is a tour for evaluators and architects: *what does Qualcomm + ROS give you for robotics, and how do I try each piece?*
```mermaid theme={null}
flowchart LR
EVAL(["Qualcomm + ROS
for robotics"])
EVAL --> NPU["NPU inference
samples · nn_inference · npu-workflows"]
EVAL --> ZC["Zero-copy data plane
qrb_ros_transport"]
EVAL --> VAL["Validation tooling
benchmark · DDS optimizer"]
EVAL --> SIM["Try without hardware
qrb_ros_simulation"]
classDef hero fill:#31017D,stroke:#31017D,color:#fff,stroke-width:1.5px;
classDef pkg fill:#F4EFFA,stroke:#31017D,color:#31017D,stroke-width:1.5px;
class EVAL hero;
class NPU,ZC,VAL,SIM pkg;
```
Load a `.tflite` / `.so` / `.bin` model and run it on the Hexagon HTP NPU from a ROS 2 node.
DMA-buf fd passing instead of payload memcpy. `Image` and `PointCloud2` types.
Working AI (detection, segmentation, pose, depth) and robotics (Nav2, AprilTag, pick-and-place) pipelines.
Evaluate QRB ROS without buying a board. Pre-built AMR and manipulator configs.
Measure zero-copy wins apples-to-apples against stock `sensor_msgs`.
LLM-driven FastDDS QoS tuner — works on any ROS 2 setup, no Qualcomm hardware required.
## How this differs from stock ROS 2
* **Hexagon HTP NPU access.** Stock TFLite / ONNX ROS nodes run inference on CPU or OpenCL GPU only. Targeting the HTP NPU requires the Qualcomm **QNN delegate** (or QNN SDK) — there is no community equivalent.
* **DMA-buf camera sharing.** Stock `image_transport` + `sensor_msgs::Image` always serialize and memcpy the payload between nodes. [`qrb_ros_transport`](./qrb-ros-transport) passes a DMA-buf file descriptor instead, so the frame the camera ISP wrote stays where it was until the next hardware consumer reads it.
* **Packaged accelerator-correct pipelines.** [`qrb_ros_samples`](./qrb-ros-samples) bundles camera ingest, QNN inference, and pre/post‑processing into launch files so you don't have to wire the plumbing yourself.
## Choosing your level of abstraction
You don't have to pick "QRB samples" or "stock ROS" — the primitives compose. Pick the level of wiring that fits your goal:
| Level | What you use | Example in this section |
| ---------------------- | ------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------- |
| **From scratch** | QNN delegate + stock ROS 2 packages + AI Hub model, wired yourself | [NPU Workflows](./npu-workflows) — the hand‑rolled depth pipeline |
| **Generic NPU node** | [`qrb_ros_nn_inference`](./qrb-ros-nn-inference) + your own model from AI Hub | Drop-in QNN inference on any topic |
| **Reference pipeline** | [`qrb_ros_samples`](./qrb-ros-samples) catalog entry, as‑is or lightly modified | [`sample_depth_estimation`](https://github.com/qualcomm-qrb-ros/qrb_ros_samples/tree/main/ai_vision/sample_depth_estimation) |
All three are valid production paths. They target the same silicon and can interoperate on the same ROS graph.
## Repositories
| Repo | Purpose | License |
| ------------------------------------------------------------------------------------------------ | ---------------------------------------- | ------------ |
| [`qrb_ros_transport`](https://github.com/qualcomm-qrb-ros/qrb_ros_transport) | Zero-copy ROS 2 transport types | BSD‑3‑Clause |
| [`qrb_ros_nn_inference`](https://github.com/qualcomm-qrb-ros/qrb_ros_nn_inference) | Generic QNN inference ROS node | BSD‑3‑Clause |
| [`qrb_ros_samples`](https://github.com/qualcomm-qrb-ros/qrb_ros_samples) | AI + robotics sample pipelines | BSD‑3‑Clause |
| [`qrb_ros_simulation`](https://github.com/qualcomm-qrb-ros/qrb_ros_simulation) | Gazebo sim of QRB AMRs and arms | BSD‑3‑Clause |
| [`qrb_ros_benchmark`](https://github.com/qualcomm-qrb-ros/qrb_ros_benchmark) | `ros2_benchmark` extension for QRB types | BSD‑3‑Clause |
| [`ROS2-DDSConfig-Optimizer`](https://github.com/qualcomm-qrb-ros/ROS2-DDSConfig-Optimizer) | LLM-driven FastDDS auto-tuner | BSD‑3‑Clause |
| [`qrb_ros_tensor_process`](https://github.com/qualcomm-qrb-ros/qrb_ros_tensor_process) | YOLO pre/post-processing nodes | BSD‑3‑Clause |
| [`qrb_ros_color_space_convert`](https://github.com/qualcomm-qrb-ros/qrb_ros_color_space_convert) | NV12 ↔ RGB8 GPU-accelerated converter | BSD‑3‑Clause |
| [`qrb_ros_image_resize`](https://github.com/qualcomm-qrb-ros/qrb_ros_image_resize) | EVA-accelerated NV12 downscaler | BSD‑3‑Clause |
The full org is at [github.com/qualcomm-qrb-ros](https://github.com/qualcomm-qrb-ros) — it includes additional packages tied to specific robot chassis, sensor subsystems, or device services that aren't highlighted here.
The QRB ROS org also contributes upstream packages that work on any Linux SoC with a DMA heap — [`dmabuf_transport`](https://github.com/qualcomm-qrb-ros/dmabuf_transport) (portable REP 2007 adapted types) and [`lib_mem_dmabuf`](https://github.com/qualcomm-qrb-ros/lib_mem_dmabuf) (userspace DMA-buf helper). `qrb_ros_transport` is built on top of both, so you don't need to touch them directly — the pointers are here for readers who want the full stack.
QRB ROS targets **Qualcomm Ubuntu for IoT** with **ROS 2 Jazzy**. The packages are distributed through the Qualcomm IoT PPAs (`ppa:ubuntu-qcom-iot/qcom-ppa` and `ppa:ubuntu-qcom-iot/qirp`). For Qualcomm Linux, refer to the [QIRP SDK](https://docs.qualcomm.com/bundle/publicresource/topics/80-70018-265/introduction_1.html?vproduct=1601111740013072\&version=1.4\&facet=Qualcomm%20Intelligent%20Robotics%20Product%20\(QIRP\)%20SDK) documentation instead.