NVIDIA Isaac SDK
Build and deploy commercial-grade, AI-powered robots. The NVIDIA Isaac SDK? is a toolkit that includes building blocks and tools that accelerate robot developments that require the increased perception and navigation features enabled by AI.Join now
The SDK features GPU-accelerated algorithms and deep neural networks (DNNs) for perception and planning, and machine learning workflows for supervised and reinforcement learning.
Modular robotic algorithms provide sensing, planning, or actuation for both navigation use cases.
Training and continuous testing in high- fidelity physics and photorealistic simulation accelerates robot development and deployment.
Rich Software Platform for AI-based Robot Development
The SDK includes the Isaac Engine (an application framework), Isaac GEMS (packages with high-performance robotics algorithms), Isaac Apps (reference applications) and NVIDIA Isaac Sim? (a powerful simulation platform). These tools and APIs accelerate robot development by making it easier to add for perception and navigation.
Delivering GPU-Accelerated Real-Time Performance Using Isaac SDK
DNN-Based GEMS Performance
|Inference Resolution||Jetson Xavier? NX Frames per Second||Jetson AGX Xavier? Frames per Second|
|13D pose: Pose convolutional neural network (CNN)|
|Object Detection (Detect_netv2, resnet18)||640x368||120||205|
|Freespace Segmentation (Unet)||512x256||49||86|
Non-DNN GEMS Performance
|Resolution||Jetson Xavier? NX Execution Time||Jetson AGX Xavier? Execution Time|
|Superpixel||640 x 360||20.72 ms||16.89 ms|
|AprilTag detection||701 x 935||17.51 ms||11.64 ms|
|2Two-lidar initial/startup global localization||1250 x 500 (binary map)||6.93 s||4.57 s|
|3Evidential grid maps (EGM) fusion||256 x 256 (local map)||8.94 ms||3.33 ms|
Featured GPU-Accelerated GEMS
GEMS are high-performance robotics algorithms.
Visual SLAM based Localization
Isaac’s VSLAM localization (preview release) offers best in class vision-based localization performance. In fact, it is a top performer when looking at real-time stereo VSLAM implementations based on publicly available results using the KITTI database. Developers interested in exploring the benefits of Visual based localization can get started in Isaac Sim or on the NVIDIA Carter robot.
Isaac Remote Control (RC)
Isaac Remote Control (preview release) provides low-latency, video-streaming for control applications to the remote robot operator. Based on Cloud XR, Isaac RC is a highly performant software stack that offers differentiated Quality-Of-Service (QOS).
Indoor Mobile Robot, Carter v2.0
Carter v2.0 is the indoor mobile robot reference design platform for Isaac SDK users. It is based on Segway Robotics’ RMPLite 220 Drivetrain with integrated IMU and is very flexible allowing for additional sensors to be added and evaluated.
Isaac Sim, Robotics Simulator
Testing and training in simulation can save time and effort. Isaac Sim provides a photo-realistic, physics accurate simulation environment that runs seamlessly with Isaac SDK. Many features in the SDK are already supported in Isaac Sim including Carter 2.0, Isaac RC, and VSLAM-based localization.
Isaac SDK Partners
Isaac SDK partners offer drivers that seamlessly integrate with the Isaac SDK. A complete list of drivers and Isaac compatible hardware can be found here. A complete view of NVIDIA’s autonomous machines ecosystem can be found here.
“BMW Group is committed to the Power of Choice for our customers—customization of diverse features across diverse vehicles for diverse customers. Manufacturing high-quality, highly customized cars, on multiple models, with higher volume, on one factory line, requires advanced computing solutions from end to end. Our collaboration with NVIDIA allows us to develop the future-of-factory logistics today and to ultimately delight BMW Group customers worldwide.”
— BMW Group
“The Government Technology Agency of Singapore developed add-on capabilities for Boston Dynamics’ ‘Spot’ legged robot that enabled us to deploy it to support safe distancing operations in a public park in Singapore. Our project included AI-based algorithms such as Visual SLAM running on the NVIDIA Jetson platform to increase the robot’s autonomy and its ability to adapt to new environments; we also leveraged the NVIDIA Isaac SDK for autonomous navigation of the ‘Spot’ robot.”
— Government Technology Agency of Singapore
Detailed View of Isaac Platform
The Isaac Engine is a software framework to easily build modular robotics applications. It enables high-performance data processing and deep learning for intelligent robots. Robotics applications developed on Isaac Robot Engine can seamlessly run on edge devices like the NVIDIA? Jetson AGX Xavier? and NVIDIA? Jetson Nano?, as well as on a workstation with a discrete NVIDIA? GPU like T4.
Compute Graph & Entity Component Architecture
- Helps break down a complex robotics use case into smaller building blocks and customizes it by configuring pre-packaged components
- Avoids host-device memory copies and increases application performance by attaching CUDA? buffer objects to messages
- Groups nodes into subgraphs, effectively combining them into a robotics application
Tools for visualization, record, replay and more
- Comes with a customizable visualization framework for creating visualizations of variable plots, drawings, camera image overlays, or 3D scenes
- Comes with Isaac WebSight for inspecting and debugging robotics applications in a web browser
- Makes it possible to writes fully functional Isaac codelets in Python
- Leverages Isaac C++ and high-performance GEMS in Python applications
- Provides easy management of Isaac applications from a Jupyter Notebook
GEMS are modular capabilities for sensing, planning, or actuation that can be easily plugged into a robotics application. For example, developers can add obstacle detection, stereo depth estimation, or human speech recognition to enrich their robot use cases. Similarly, developers can use the Isaac navigation stack, a collection of modules, to build a robotics navigation application for their robot.
Deep object pose estimation (DOPE) provides a novel model to estimate the six degrees of freedom (6DOF) pose of an object using an RGB image. It’s suitable for manipulation tasks.
An application that allows users to submit missions to a robot and control multiple robots via programmable mission dependencies
A planning algorithm that avoids obstacles by checking trajectories for collisions, range, and speed before execution
A customizable global planning algorithm that leverages semantics zone information like reduced speed, restricted access, or unidirectional movement
An algorithm that provides for smooth, predictable, and straight forward and backward movement
A flexible interface for LQR planner cost functions, which can be customized, combined and used on different robots
A well defined process to help Jetson depth camera partners bring up their cameras on the Isaac SDK
An application that demonstrates the breadth of the Isaac SDK’s navigation and manipulation capabilities, showing how robots work together to complete a mission in a simulated factory environment.
An application that shows our Isaac SDK pick-and-place manipulation capabilities using the UR10 robot arm.
An NGC? docker that makes it easy to experiment in simulation (Unity 3D) with the Isaac Pose CNN GEM
Follows a ball in simulation to demonstrate how to do training and inference in simulation for a low-cost robot
Isaac Sim provides the essential features for building virtual robotic worlds and experiments. Isaac Sim supports navigation and manipulation applications through the Isaac SDK with RGB-D, lidar and inertial measurement unit (IMU) sensors, domain randomization, ground truth labeling, segmentation, and bounding boxes.
Training of pose estimation pipeline using machine learning workflows possible with Isaac SDK
Shows random changes in material (texture, color), light direction, light conditions, sunlight changes, placement of objects/obstacles, floors, etc., to train robot perception and behavior as well as test for robust behavior in real life
Multiple Carter robots operating simultaneously in a virtual warehouse. Each operated by an independent Jetson Xavier
Build and deploy commercial-grade, AI-powered robots using the Isaac SDK.