According to reports from Electronic Enthusiasts (by Wu Zipeng), in 2025, robotics technology is accelerating its transition from the era of specialized machines to that of general-purpose robots. This process not only signifies a vast expansion in the application range of robots but also indicates that robots will play a crucial role in more industries. The “White Paper on the Development of Humanoid Robots in China (2025)” shows that 2025 is hailed as the “Year of Mass Production” for humanoid robots, with shipments expected to exceed 20,000 units and a market scale approaching 9 billion yuan; it is anticipated that by 2035, global demand for humanoid robots will reach 18 million units, with the market scale climbing to 1.5 trillion yuan.In this wave of transformation, NVIDIA’s Jetson AGX Thor developer kit and Jetson T5000 module act as a “shot in the arm,” injecting powerful momentum into the global developer community to build a physical AI future and pushing the development of general robotics into a new phase.Deepu Talla, Vice President of Robotics and Edge Computing at NVIDIA, stated: “We do not manufacture robots, but we empower the entire industry through our computing platform. Currently, several humanoid robot companies in China have taken the lead in upgrading to NVIDIA Blackwell-based Jetson Thor, including UBTECH, Galaxy Universal, Yushu Technology, Zhongqing Robotics, and Zhiyuan Robotics.”It is reported that Jetson Thor is the core system of the Jetson AGX Thor developer kit, featuring 128GB of memory and AI computing performance of up to 2070 FP4 TFLOPS, capable of easily running the latest generative AI models, all within a power range of 130 watts. Therefore, compared to the previous generation NVIDIA Jetson Orin, Jetson Thor’s AI computing performance has increased by 7.5 times, CPU performance by 3.1 times, and memory capacity by 2 times, meeting real-time processing needs at the edge. Additionally, Jetson Thor introduces native FP4 quantization technology and a next-generation Transformer engine, which can dynamically switch between FP4 and FP8 precision, combining 4-bit weights, activation values, and higher memory bandwidth to accelerate the pre-filling and decoding processes of generative AI workloads simultaneously.With these advantages, Jetson Thor can smoothly run various generative AI models. Through optimization of the Jetson software stack, this platform meets the dual demands of low latency and high performance for real-time applications, supporting all mainstream generative AI frameworks and AI inference models, with significant real-time performance advantages. These models include general models such as Cosmos Reason, DeepSeek, Llama, Gemini, Qwen, as well as robot-specific models like Isaac GR00T N1.5, allowing developers to easily conduct model experiments and inference operations locally.Building a truly general humanoid robot requires constructing a complex and precise system architecture, typically divided into four core levels:
- Hardware Abstraction Layer: Integrating key sensors such as cameras, IMUs (Inertial Measurement Units), and LiDAR, as well as actuators like motors, enabling the robot to perceive its environment and interact with the physical world.
- Real-Time Control Framework Layer: Providing precise, low-latency control of robot movements to ensure safe operation and rapid response.
- Perception and Planning Layer: Endowing the robot with capabilities for environmental understanding, object recognition, path navigation, and motion planning.
- High-Level Reasoning Layer: Supporting advanced functionalities such as scene understanding, complex task planning, and natural language interaction..
The launch of the Jetson AGX Thor developer kit and Jetson T5000 module can help designers effectively address key issues in building the aforementioned architecture. Among them, the NVIDIA Jetson AGX Thor developer kit boasts unparalleled performance and scalability, relying on the new Blackwell Multi-Instance GPU (MIG) and a powerful 14-core Arm Neoverse-V3AE CPU to accelerate low-latency, high-real-time applications. Furthermore, this platform integrates a complete set of accelerators, including the third-generation Programmable Vision Accelerator (PVA), dual encoders and decoders, and optical flow accelerators.In terms of high-speed sensor fusion, the developer kit offers a rich array of input/output (I/O) interface options, including 1 QSFP slot supporting 4 channels of 25GbE, 1 wired Multi-GbE RJ45 interface, multiple USB ports, and other connection interfaces. Its design allows for seamless integration with existing humanoid robot platforms, supporting convenient wired connections to help developers quickly initiate prototype development.On the software side, the Jetson software accelerates edge AI performance through a highly integrated full-stack software platform. This platform is driven by JetPack 7, featuring a Linux 6.8 kernel, Ubuntu 24.04 LTS operating system, and the latest NVIDIA AI stack, enabling low-latency, high-deterministic operation of advanced generative AI models, providing stable support for physical AI applications. Integrated features such as the Holoscan Sensor Bridge, MIG support, and Preemptable Real-Time Kernel significantly enhance the performance and efficiency of tasks like high-speed sensor fusion and motion planning. Notably, the Jetson Thor platform supports the new Cosmos Reason—a customizable, open-source, 7 billion parameter inference-based visual language model (VLM) aimed at the physical AI and robotics fields. Additionally, with JetPack 7, the Jetson software aligns with the Server Base System Architecture (SBSA), allowing Jetson Thor to meet industry standards for Arm server design, simplifying the development process, reducing version fragmentation issues, and ensuring consistency in experience from server-level systems to the Jetson Thor platform.Currently, the NVIDIA Jetson AGX Thor developer kit (priced at $3499) and the Jetson T5000 production-level module are officially available through NVIDIA’s global authorized distributors. Beyond the humanoid robotics field, Jetson Thor will also accelerate various robotic applications such as surgical assistance robots, smart towing vehicles, delivery robots, industrial robotic arms, and visual AI agents, providing real-time inference capabilities for larger and more complex AI models at the edge.
Disclaimer：This article is originally from Electronic Enthusiasts, please indicate the source above when reprinting. For group discussions, please add WeChat elecfans999, for submission and interview requests, please email [email protected].
More Hot Articles to Read
-
AI Year, Sensor Carnival! Three Major Sensor Companies H1: Highest Revenue Exceeds 3.7 Billion, One Achieves Profit for the First Time
-
Huawei Cloud Architecture Major Adjustment, Affecting Thousands of Employees, Can All in AI Turn Losses into Profits?
-
From GPU to ASIC, AI Server Power Chips Explode!
-
Global Humanoid Robot’s First 10,000 Unit Order! Surprisingly Not UBTECH or Yushu?
-
DeepSeek V3.1 Released! Embracing Domestic Computing Power Chips
Click to Follow and Bookmark Us
Set us as a bookmark to not miss any updates!