General-purpose robots and autonomous machines are transitioning from research labs to widespread real-world applications, highlighting the demand for compact and power-efficient AI supercomputers that can effectively run models at the edge.
In response to this emerging need, NVIDIA has unveiled NVIDIA Thor— an innovative architecture designed to support mass-market robotics and edge AI applications at scale.
Jetson AGX Thor is powering the next generation of humanoid and robotic systems, experiencing rapid adoption across various industries. Leading companies—including 1X, Agile Robot, Amazon Robotics, Boston Dynamics, Fanuc, Hitachi, and Techman Robot— are leveraging this platform.
Facilitating Humanoid and Robot Deployment with T3000
The foundation for these advanced capabilities lies in the Jetson and IGX T3000 modules, offering a stunning 865 FP4 teraflops of AI computing in a compact design that consumes about half the size and power of the T5000. The Jetson T3000 integrates an NVIDIA Blackwell GPU, an 8-core Neoverse Arm CPU, 32 GB of LPDDR5X memory, 273 GB/s of memory bandwidth, and 25 GbE connectivity. Meanwhile, the IGX T3000 delivers the same performance with built-in functional safety, alongside NVIDIA Halo for Robotics— a comprehensive safety system for collaborative robots.
Despite its smaller footprint, the T3000 matches the T5000 in inference performance for multimodal workloads, including large language models, vision-language models, and world infrastructure models. Transitioning to the T3000 helps cut costs amidst rising memory prices.
Broad Applications of Edge AI with T2000
Jetson T2000 extends the Thor architecture to a wider array of edge AI systems. With 400 FP4 teraflops of compute power and 16 GB of memory, it serves as an ideal entry point for developers creating visual AI agents, autonomous mobile robots, industrial manipulators, and intelligent machines.
The introduction of the NVIDIA Jetson module signifies a scalable edge AI platform ranging from 70 TOPS to 2,000 teraflops of performance, empowering developers to tackle virtually any edge AI workload.
New Agent Skill Enhances Memory Optimization for All Jetson Devices
AI agents are revolutionizing developer productivity by automating essential tasks like memory optimization, system configuration, and deployment processes that traditionally required extensive manual effort and expertise.
The newly launched Jetson Agent Skills enable developers to optimize their entire software stack, achieving considerable memory savings in days rather than weeks. This feature supports the complete Jetson portfolio, including Jetson Thor and Jetson Orin, allowing developers to execute higher-performance workloads in lower memory configurations.
The outcome is decreased system costs and accelerated deployment, with the flexibility to utilize lower memory SKUs within the same product tier without sacrificing performance.
Organizations across various sectors are speeding up development while realizing substantial memory savings through optimized software.
A leading player in humanoid robotics, Ubitech, and Agile Robot, in collaboration with industrial solution providers like Connect Tech, have reduced memory usage by up to 15 GB, enabling migration from NVIDIA Jetson AGX Orin 64GB to 32GB modules.
In smart retail, Sand Star has achieved a memory usage reduction of up to 4GB, facilitating deployment with NVIDIA Jetson Orin NX 8GB modules instead of 16GB configurations. In companion robotics, Groove X, the creators of the LOVOT robot, leverage Jetson’s heterogeneous AI accelerators to optimize workload distribution and enable deployment in low-memory settings.
In intelligent transportation, Traffic Prohibited has reduced memory usage by 30% with Jetson TX2 NX, allowing for the addition of AI capabilities to their smart traffic platform without increasing hardware requirements.
Jetson serves as an agent-enabled platform for physical AI, offering agent skills to simplify development, while NVIDIA NemoClaw blueprints orchestrate intelligent agents for advanced inference, autonomous decision-making, and task automation on a large scale.
Introducing Cosmos 3 Edge to the NVIDIA Thor Lineup
NVIDIA has also announced NVIDIA Cosmos 3— a groundbreaking open-world foundation model specifically designed for robotics. Cosmos 3 Edge features lightweight models compatible with the NVIDIA Thor platform. This 4 billion parameter model assists embodied systems in perceiving their environment, reasoning in real-time, and generating actions via on-device inference. The open Cosmos framework allows developers to fine-tune Cosmos 3 Edge for specific embodiments and sensors in about 1 day— bridging the gap between simulation and reality and enabling the deployment of Jetson Thor for real-time vision analytics and on-device robot policy.
Begin Developing Today in Emulation Mode
Sharing the same chip architecture and software stack as the NVIDIA Thor family, the new module provides a straightforward development path. Developers can start building now with the Jetson AGX Thor developer kit available below. Channel partners offer emulation of the performance of T3000 and T2000 modules.
Utilize all NVIDIA features with Physics AI alongside our software stack, which includes NVIDIA Isaac for robotics simulation and recognition, as well as open models like NVIDIA Nemotron Cosmos 3, and Isaac GR00T to accelerate the development of next-generation robots, autonomous machines, and visual AI agents.
Developers can begin utilizing T3000 emulation mode in JetPack 7.2.1 later this month. Support for T2000 emulation mode will be added in a future release. The Jetson T3000 and T2000 modules are anticipated to be available in Q1 2027.
Key industry partners, including ADLINK, Advantech, Ion, Aetina, Aubidea, Avermedia, Connect Tech, Fore CR, JWIPC, NEXCOM Robot Solution, Real Time, Seeed Studio, Two Win, Tutek, and Former are also partnering to deliver Thor-based solutions providing emulation and migration support for customers transitioning to the new modules.
As physical and embodied AI gain traction, the new NVIDIA Thor computers equip developers with a scalable foundation to launch intelligent humanoids and autonomous machines into the real world.
Get started with the Jetson AGX Thor developer kit at the NVIDIA Marketplace and begin your development journey today.
Source: blogs.nvidia.com