When robots perform precise stacking in warehouses, assist doctors during surgeries, or complete search and rescue missions in complex environments, they rely on a powerful ‘brain’.On August 25, NVIDIA announced that its latest robotic chip module, the Jetson AGX Thor, is now available as a developer kit for $3,499.NVIDIA refers to this chip as the ‘robot brain’.
Performance Explosion: From ‘Slow Response’ to ‘Real-Time Response’When robots operate in the real world, they must simultaneously process data streams from multiple sensors such as cameras, radar, and LiDAR, while quickly completing AI inference and decision-making, which requires extremely high computational power.Previously, many complex robotic tasks suffered from insufficient hardware performance, resulting in either ‘slow responses’ or reliance on cloud computing, making it difficult to adapt to dynamic physical environments.The introduction of Jetson Thor directly breaks this limitation.Compared to the previous generation product, Jetson Orin, it achieves a comprehensive leap in core performance:– AI computing performance increased by 7.5 times:up to 2070 FP4 TFLOPS, easily handling complex AI tasks such as large Transformer models and visual language models– CPU performance increased by 3.1 times:multi-tasking is smoother, and the robot’s ‘thinking’ pace is faster– Memory capacity doubled:able to cache more sensor data and AI models simultaneously, avoiding task interruptions due to insufficient memoryMore importantly, Jetson Thor has undergone extensive optimizations for ‘real-time performance’.Through FP4 precision computing and inference decoding technology, it enables AI models to achieve ‘instant response’ at the edge (i.e., locally on the robot) — for example, outputs that previously took 50 milliseconds to generate can now be produced much more efficiently, even running AI workloads that were previously only feasible on servers, minimizing reliance on the cloud.
Not Just Humanoid Robots: Covering All Scenarios of Robotic ApplicationsWhen people think of robots, many envision humanoid devices like ‘Transformers’, but Jetson Thor has already become the ‘core choice’ for leading humanoid robot companies.Leading company Agility Robotics’ fifth-generation commercial robot, Digit, is already handling stacking and loading tasks in warehouses, and they plan to equip the sixth-generation Digit with Jetson Thor.Its CEO, Peggy Johnson, stated: ‘Jetson Thor enhances Digit’s real-time response capabilities and expands its ability to tackle more complex tasks, helping customers improve operational efficiency in warehouses and factories.’Another industry benchmark, Boston Dynamics, is also integrating Jetson Thor into its humanoid robot, Atlas.This means Atlas will have server-level computing power on-device for the first time, easily handling high-bandwidth sensor data processing and large memory support, potentially enabling more flexible actions such as running, jumping, and grasping in the future.In addition to humanoid robots, Jetson Thor is also making significant contributions in various fields:– Medical field: Surgical assistance robots can process multi-camera data in real-time, providing precise guidance to doctors– Industrial scenarios: Intelligent towing vehicles and industrial robotic arms can operate efficiently in unstructured environments– Public services: Delivery robots can quickly recognize road conditions, and visual AI agents can monitor factory worker safety…It can be said that any robot requiring ‘perception – decision – action’ can become smarter with the addition of Jetson Thor.
Developer-Friendly: Comprehensive Software and Hardware EcosystemAn excellent hardware platform is supported by a robust ecosystem.Jetson Thor not only boasts strong performance but also provides developers with a complete toolchain for ‘from R&D to mass production’, significantly lowering the innovation threshold.On the software side, it is compatible with all mainstream generative AI frameworks, supporting general models such as Cosmos Reason, Llama, Gemini, and robot-specific models like Isaac GR00T N1.5 — developers do not need to build models from scratch and can conduct experiments and inference locally.Additionally, it can run the complete NVIDIA AI software stack:– Use NVIDIA Isaac for robot motion control and path planning– Use NVIDIA Metropolis for video analysis, creating visual AI agents– Use NVIDIA Holoscan to simplify sensor fusion, enabling ultra-low latency transmission of camera and radar data to the GPUIn terms of hardware, Jetson Thor offers two options: The developer kit includes the T5000 module, a reference carrier board with rich interfaces, an active cooler, and power supply, making it easy to get started; the production-grade module is suitable for bulk purchases by enterprises, priced at $2,999 for orders of 1,000 units or more.Moreover, partners such as Advantech, Miwen Power, and Tianzhun Technology are developing customized production systems to meet different interface and form factor requirements.Even in the research field, Jetson Thor has become highly sought after.Laboratories at universities such as Stanford and Carnegie Mellon are using it to push the boundaries of perception and navigation models — the Carnegie Mellon team plans to upgrade their existing systems to Jetson Thor to enhance the medical triage and search and rescue capabilities of autonomous robots, and even conduct experiments with robot swarms.
How to Get Started? Prices and Activities HereCurrently, the NVIDIA Jetson AGX Thor developer kit is officially on sale for $3,499; the Jetson T5000 production module is also available for purchase, priced at $2,999 for bulk orders (1,000 units or more), which can be ordered through NVIDIA authorized partners.If you want to experience the capabilities of Jetson Thor firsthand, you can also sign up for the hackathon event jointly initiated by Seeed Studio and Hugging Face’s LeRobot, exploring more possibilities of physical AI with developers worldwide.
From laboratories to factory workshops, from warehouses to operating rooms, the emergence of Jetson Thor is turning the concept of ‘smarter and more flexible robots’ into reality.Jetson Thor unlocks real-time inference capabilities for general-purpose robots and physical AI.When robots can process complex environments in real-time and autonomously complete challenging tasks, it may not be long before we see these ‘intelligent partners’ in more scenarios.In the future, more possibilities of physical AI will be unlocked one by one.Text Editor:imoo.o Image Source:NVIDIAReferences:NVIDIASpecial Statement:If there are any errors in source attribution or infringement, please contact us promptly, and we will correct or delete it immediately.If you want to chat more about AI with imoo.o 
click on the digital avatar below.