Reported by Electronic Enthusiasts (Written by Wu Zipeng), recently, embodied intelligence has become a common topic at various artificial intelligence-related conferences, and many people define 2024 as the first year of the embodied intelligence industry development, as previous explorations were mostly conceptual, while this year marks the beginning of actual demand implementation.

According to the definition by the China Computer Federation, embodied intelligence spans multiple disciplines including artificial intelligence, robotic learning, and computer vision, representing a research paradigm where intelligent agents with physical bodies interact with the physical environment to gain intelligence. So, how can engineers build their own embodied intelligence innovative solutions?

Here, we recommend NVIDIA’s Jetson Thor chip solution and its related resources.

The Jetson Thor, designed specifically for robotic applications

In 1950, Turing first proposed the concept of embodied intelligence, but after that, significant progress was not made, largely due to the limitations of the software and hardware resources at that time, making embodied intelligence seem like a fantasy. At the ITF World 2023 Semiconductor Conference, NVIDIA’s founder and CEO Jensen Huang reiterated the concept of embodied intelligence, believing that the next wave of artificial intelligence will be embodied intelligence.

At GTC 2024, NVIDIA released the Jetson Thor chip platform specifically designed for humanoid robots. This chip incorporates a wealth of NVIDIA’s know-how regarding the development of embodied intelligence. Before the release of Jetson Thor, NVIDIA had already launched several chip solutions in the robotics field, including Jetson Orin, Jetson Orin Nano, and Jetson AGX Xavier.

Jetson hardware roadmap, image source: NVIDIA

Furthermore, as NVIDIA stated, each Jetson series product is a complete System on Module (SOM), which includes GPU, CPU, memory, power management, and high-speed interfaces. Taking the Jetson Orin module as an example, it can provide up to 275 TOPS of processing power, with memory configurations ranging from 4GB to 64GB, and module power consumption between 7 to 60 watts.

Jetson Thor continues the high-performance tradition of the Jetson series, providing 800 TOPS of 8-bit floating-point AI performance, and integrates a functional safety processor, high-performance CPU clusters, and 100GB Ethernet bandwidth. As mentioned in my previous article, Jetson Thor is based on NVIDIA’s Blackwell architecture, ensuring high performance for the chip. Compared to NVIDIA’s previous Hopper architecture, the Blackwell architecture brings several times of improvement in various performance aspects. Additionally, the Blackwell Tensor core adds new precision, supporting 4-bit floating-point AI inference, effectively doubling computing capability and model size.

According to NVIDIA’s introduction, Jetson Thor can execute complex tasks and interact safely and naturally with humans and machines, featuring a modular architecture optimized for performance, power consumption, and size.

Rich development resources around the Jetson series

As mentioned above, the NVIDIA Jetson series is a solution for robotic development that meets performance and budget needs for various applications. In this solution, besides the high-performance, highly integrated chip platform, there are rich supporting resources.

For product developers, the Jetson series integrates the JetPack SDK, Metropolis microservices, production-ready Isaac ROS packages, and application-specific reference AI workflows.

The JetPack SDK is the source of the Jetson software stack, providing Jetson Linux, developer tools, CUDA-X acceleration libraries, and other NVIDIA technologies. The JetPack SDK builds the underlying capabilities for various robotic applications, allowing developers to use TensorRT and cuDNN for AI inference acceleration, CUDA for general computing acceleration, VPI for computer vision and image processing acceleration, and Jetson Linux API for multimedia acceleration, with libArgus and V4l2 for camera processing acceleration.

At the same time, with Jetson Linux, developers can develop various applications, as the toolkit includes the Linux kernel, bootloader, NVIDIA drivers, flashing utilities, example file systems, and toolchains for the Jetson platform, making application development very swift. In summary, the JetPack SDK contains a wealth of segmented functional SDKs, allowing engineers to choose based on their needs.

Next, let’s look at the Isaac ROS package, which is a hardware-accelerated package for robotic operating systems, including multiple open-source options. Isaac ROS provides a single software package (GEM) with image processing and computer vision capabilities, and a complete pipeline (NITROS), helping robotic applications build high-throughput perception systems; the modular software of Isaac ROS allows developers to replace algorithms as needed; Isaac ROS has excellent compatibility, significantly shortening the development cycle of robotic solutions. In summary, Isaac ROS is a rich and modular perception software package, providing high-performance perception and hardware acceleration for robotic applications.

For educators, students, and enthusiasts, NVIDIA also offers the Jetson Nano Developer Kit, which helps these developers teach, learn, and develop AI and robotics integration solutions.

In addition to the rich supporting development resources, we must also reiterate NVIDIA’s advanced concepts in large models for embodied intelligence. The NVIDIA Project GR00T humanoid robot base model is a VLA-type model that possesses language, action, and specialized robotic knowledge modalities, representing the mainstream model route for future embodied intelligence development. Developers can access these knowledge and resources from the NVIDIA GR00T project or Isaac toolkit.

Conclusion

The NVIDIA Jetson Thor chip is a high-performance inference chip specifically designed for robotic applications, and it provides a hardware solution for future embodied intelligence applications within the Jetson series. High computing power, high integration, and an all-in-one solution are the characteristics of this platform.

At the same time, NVIDIA also provides rich software resources, from the lowest-level SDK to operating system-level resources, enabling rapid development of robotic applications such as embodied intelligence.

Disclaimer: This article is original by Electronic Enthusiasts, please indicate the source above when reprinting. For group discussions, please add WeChat elecfans999,for submission or interview requests, please send an email to [email protected].

More Hot Articles to Read

• Zhijie S7, Zhiji L6 Promoting the “Smart Chassis” Concept, What New Opportunities Are There in the Industry Chain?

• DPU Market at a New Crossroads, Is Third-Party Not as Good as Self-Development?

• Power Device Manufacturers IGBT Revenue Comparison: Some Significantly Stockpiling and Expanding Production, Some Supply Exceeds Demand and Plummets

• Huawei Re-releases Zhijie S7! Multiple Versions Reduced by 20,000! P70 Ignited Before Release, 980g Ultra-Light Laptop Debuts

• Contemporary Amperex Technology Co. Ltd. Releases Tianheng Energy Storage System, Achieving 5 Years Zero Degradation, Power Battery Manufacturers Gather for Energy Storage

Click to Follow and Bookmark Us

Bookmark us so you don’t miss any updates!

If you like it, reward us with a “look”!