NVIDIA Isaac ROS

Latest Humble ROS 2 release improves performance on compute platforms that offer hardware accelerators. Humble enables hardware-acceleration features for type adaptation and type negotiation eliminating software/CPU overhead and improving performance of hardware acceleration.
The NVIDIA implementation of type adaption and negotiation is called NITROS. These are ROS processing pipelines made up of Isaac ROS hardware accelerated modules (a.k.a. GEMs) and added to NVIDIA’s latest Isaac ROS Developer Preview (DP) release.

NVIDIA Isaac Transport for ROS

H.264 video encode and decode hardware-accelerated packages for NITROS, used for compressed camera data recording and playback for development of AI models and perception functions, compresses 2x 1080p stereo cameras at 30fps (>120fps total) and reduces the data footprint by ~10x.

As autonomous machines move around in their environments they must keep track of where they are. Visual odometry solves this problem by estimating where a camera is relative to its starting position. The Isaac ROS GEM for Stereo Visual Odometry provides this powerful functionality to ROS developers.

This GEM offers the best accuracy for a real-time stereo camera visual odometry solution. Publicly available results based on the widely used KITTI database can be referenced here. For the KITTI benchmark, the algorithm achieves a drift of ~1% in localization and an orientation error of 0.003 degrees per meter of motion. In addition to being very accurate, this GPU accelerated package runs extremely fast . In fact, it’s now possible to run SLAM on HD resolution (1280x720) in real time (>60fps) on an NVIDIA Jetson Xavier AGX™. (>60fps) on a Jetson Xavier AGX.

Knowledge of a robot’s position alone isn't enough to safely navigate complex environments. Robots must also be able to discover obstacles on their own. nvblox (preview) uses RGB-D data to create a dense 3D representation of the robot's environment. This includes unforeseen obstacles that could cause a danger to the robot if not observed in real-time.This data helps generate a temporal costmap for navigation stack.

DNN Inference GEM is a set of ROS2 packages that allow developers to use any of NVIDIA’s numerous inference models available on NGC or even provide their own DNN. Further tuning of pre-trained models or optimizations of developers' own models can be done with the NVIDIA TAO Toolkit.

After optimization, these packages are deployed by TensorRT or Triton, NVIDIA’s inference server. Optimal inference performance will be achieved with the nodes leveraging TensorRT, NVIDIA’s high performance inference SDK. If the desired DNN model isn't supported by TensorRT then Triton can be used to deploy the model.

Additional GEMs incorporating model support are available and include support for U-Net and DOPE. The U-Net package, based on TensorRT, can be used for generating semantic segmentation masks from images. The DOPE package can be used for 3D pose estimation for all detected objects.

This tool is the fastest way to incorporate performant AI inference in a ROS application. The pre-trained model, PeopleSemSegNet, pictured in the image (right) runs at 325 fps @544p on Jetson AGX Orin.

In addition to NITROS pipelines, we also have the two new GEMs, first is ESS, which is a DNN for stereo camera disparity prediction and second one is Bi3D which is a DNN for vision-based proximity detection.

Both Bi3D and ESS are pre-trained for robotics applications using synthetic data and are intended for commercial use.

Isaac Mission Dispatch allows a cloud/edge system to send and monitor tasks from a ROS 2 robot via Isaac Mission Client using industry standards for production deployments. Mission Dispatch is a cloud-native microservice that can be integrated as part of larger fleet management systems.

Mission Dispatch and Mission Client are both available in open source and can be used to test robots in simulation for automating test portions of continuous integration and continuous deployments (CI/CD), performing a series of predefined tasks evaluated against expected results. This benefit is in addition to the primary usage of assigning tasks to robots in operation.

Mission Dispatch can be integrated into fleet management systems (e.g., Anyfleet, Roborunner FleetGateway) with Mission Client on the ROS 2 robot. Mission Dispatch will interoperate with other ROS 2 Clients built on VDA5050

In a typical robotics image processing pipeline, raw data from the camera sensor must be processed before being passed off to a DNN or classic computer vision module for perception processing. This image processing consists of things like Lens Distortion Correction (LDC), image resizing, and image format conversion. If stereo cameras are involved then estimating disparity is also required. The image processing GEMs have been designed to take advantage of the specialized computer vision hardware available in Jetson like the GPU, the VIC (Video and Image Compositor) and the PVA (Programmable Vision Accelerator).

For robots using cameras connected via a CSI interface, NVIDIA provides the hardware accelerated Argus package.

Image shows lens distorted camera image (left) and rectified image using LDC GEM (right)