Robotics technology is undergoing a transformation, gradually moving away from the era of specialized machines and stepping into the age of general-purpose robots. This shift signifies that robots are no longer confined to single-purpose, fixed-function forms but are evolving into more adaptive machines capable of performing diverse tasks across different environments. Inspired by human cognitive models, adaptive robots combine rapid response capabilities with high-level reasoning and planning abilities, enabling more efficient learning and adaptation to their surroundings.


This model opens the door to flexible applications of robotics across various industries, not only reducing costs but also expanding the practical scope far beyond the capabilities of specialized robots. At the 2025 GTC conference, NVIDIA officially launched the NVIDIA Isaac GR00T platform, laying the foundation for this transformation. The platform integrates robot foundation models, synthetic data workflows, simulation environments, and runtime computing, providing comprehensive support for the development of general-purpose robots.

The NVIDIA Jetson AGX Thor Developer Kit and NVIDIA Jetson T5000 module are now officially available. This will empower developers worldwide to shape the future of physical AI. With Jetson Thor, robots no longer require reprogramming for every new task. Jetson Thor is the premier platform for physical AI, supporting generative inference while offering multimodal, multi-sensor processing capabilities. Integrating Jetson Thor into next-generation robots accelerates foundation models, granting them greater flexibility in challenging tasks such as object manipulation, navigation, and executing complex instructions.

Pre-order Now 🚀🚀🚀

The next-generation physical AI engine—NVIDIA DRIVE AGX Thor—is now available for pre-order! Secure your competitive edge by prioritizing your order through ZENTEK. Want to learn more? For details, contact our customer service via the [Contact Customer Service] section under the ZENTEK official account, call 177-4088-9963 directly, or email us at hello@zentek.com.cn.



What Does It Take to Build a General-Purpose Humanoid Robot?



Building a typical general-purpose humanoid robot requires constructing four core layers:

• Hardware Abstraction Layer :  Integrates all key sensing and actuation modalities, enabling the robot to perceive its environment and interact physically with the world.
  
  
• Real-Time Control Framework Layer :  Ensures precise, low-latency control of robot movements. Minimizing latency is critical for safe operation and rapid responsiveness.
  
• Perception and Planning Layer :  Equips the robot with environmental understanding, grasp and motion planning, mobility, object recognition, and localization capabilities, allowing effective interaction with its surroundings. Unlike the real-time control layer, this tier can tolerate slightly longer processing times to ensure thorough and accurate decision-making.
  
  
• High-Level Reasoning Layer :  Supports advanced functionalities such as scene understanding, complex task planning, and natural language interaction. At this level, longer processing times are acceptable to achieve deeper reasoning and greater adaptability.

Figure 1. Building a humanoid robot requires integrating a range of hardware and software components.



Why Is NVIDIA Jetson Thor the Premier Platform for Physical AI and Humanoid Robots?



The NVIDIA Jetson AGX Thor Developer Kit offers unparalleled performance and scalability. Featuring an NVIDIA Blackwell architecture GPU with 128 GB of memory and AI computational performance of up to 2,070 FP4 TFLOPS, it effortlessly runs the latest generative AI models—all within a 130-watt power envelope. Compared to NVIDIA Jetson AGX Orin, it delivers 7.5x the AI computational performance and 3.5x the energy efficiency.


Leveraging the new Blackwell Multi-Instance GPU (MIG) and a powerful 14-core Arm Neoverse-V3AE CPU, Jetson Thor accelerates low-latency, real-time applications. Additionally, the platform integrates a comprehensive suite of accelerators, including a third-generation Programmable Vision Accelerator (PVA), dual encoders and decoders, and an optical flow accelerator.


For high-speed sensor fusion, the developer kit offers rich input/output (I/O) interface options, including one QSFP slot supporting 4x 25GbE, one multi-GbE RJ45 port, multiple USB ports, and other connectivity interfaces. Its design enables seamless integration with existing humanoid robot platforms, supporting wired connections for rapid prototyping.


Transformer Engine and FP4 Support

Jetson Thor, built on the NVIDIA Blackwell architecture, introduces native FP4 quantization technology paired with a next-generation Transformer Engine that dynamically switches between FP4 and FP8 precision for optimal performance. By combining 4-bit weights, activation values, and higher memory bandwidth, Jetson Thor accelerates both prefill and decoding processes in generative AI workloads.

Multi-Instance GPU (MIG)

Jetson Thor introduces MIG technology, allowing a single GPU to be partitioned into multiple isolated instances, each with dedicated resources. This ensures computational resources are reserved for critical workloads while concurrently running less time-sensitive tasks, guaranteeing predictable performance—a crucial feature for robotics applications handling multiple mission-critical tasks.

Tables 1 and 2 detail the core features and interface specifications of the Jetson T5000 module and the NVIDIA Jetson AGX Thor Developer Kit carrier board, respectively.


Table 1. NVIDIA Jetson AGX Thor Developer Kit Module Specifications

Note: Parameters marked with * are preliminary and subject to change.

Table 2. NVIDIA Jetson AGX Thor Developer Kit Carrier Board Specifications



Figure 2. Component Composition of the NVIDIA Jetson Thor Module



How Does Jetson Thor Accelerate Generative AI at the Edge?



Jetson AGX Thor is a全新的 robotics computer, redesigned from the ground up to power the next generation of humanoid robots. The platform supports a wide range of generative AI models, including vision-language-action (VLA) models like NVIDIA Isaac GR00T N1.5, as well as all major large language models (LLMs) and vision-language models (VLMs).

To provide a seamless cloud-to-edge experience, Jetson Thor runs NVIDIA's AI software stack for physical AI applications, including NVIDIA Isaac for robotics, NVIDIA Metropolis for visual AI agents, and NVIDIA Holoscan for sensor processing. Developers can also build AI agents at the edge using NVIDIA's embodied AI workflows, such as video search and summarization (VSS).




Figure 3. Jetson Thor Supports Various AI Frameworks and Generative AI Models



Why Are Generative Inference and Multimodal Sensor Processing Critical for Physical AI?



Generative inference models are essential for robot platforms that need to simulate possible action sequences, predict outcomes, reason based on language or visual cues, and flexibly generate high-level plans or low-level motion strategies. These models enable robotic systems to achieve greater flexibility, stronger adaptability, and human-level robust reasoning in real-world scenarios.

NVIDIA Jetson Thor represents a giant leap in generative inference, delivering up to 5x faster inference speeds compared to Jetson Orin. With FP4 precision optimization and speculative decoding technology, developers can achieve an additional 2x performance improvement on Jetson Thor.



Figure 4. Jetson Thor Delivers Up to 5x Faster Generative Inference Speeds Compared to Jetson Orin


Jetson Thor also seamlessly handles multiple generative AI models and extensive multimodal sensor inputs for real-time responsiveness. Figure 5 demonstrates this capability using the Qwen2.5-VL-3B VLM and Llama 3.2 3B LLM models, simultaneously processing 16 concurrent requests. The results show that both models achieve a "Time to First Token (TTFT)" well below 200 milliseconds and a "Time per Output Token (TPOT)" well below 50 milliseconds—key metrics for measuring system responsiveness.



Figure 5. Jetson Thor Maintains Real-Time Responsiveness While Handling Multiple Generative AI Models and Extensive Multimodal Sensor Inputs


Jetson Thor not only supports native FP4 precision on the Blackwell architecture but also enables advanced techniques like speculative decoding. In speculative decoding, a small draft model generates candidate tokens, which are then validated by a larger model. This approach accelerates generative AI inference while ensuring output accuracy, resulting in faster and higher-quality generation.

Figure 4 shows that with FP4 quantization and Eagle-based speculative decoding, the Qwen2.5-VL-7B model achieves up to 3.5x faster inference speeds on Jetson Thor compared to Jetson Orin using a W4A16 (4-bit weights, 16-bit activations) configuration.

Additionally, as shown in Table 3, Jetson Thor delivers breakthrough acceleration across various generative AI models, including LLMs, VLMs, and VLAs, with performance improvements of up to 5x over Jetson Orin.



Table 3. Benchmark Comparison Between Jetson Thor and Jetson AGX Orin



Benchmark configuration details: Sequence length 2048, output sequence length 128; maximum concurrency 8; LLM and VLM models run on the VLLM framework, VLA models run on the TensorRT framework; both Jetson AGX Thor and Jetson AGX Orin use MAXN power mode.


How Does Jetson Software Accelerate AI at the Edge?


Jetson software accelerates edge AI performance by providing a highly integrated full-stack software platform optimized for real-time, high-throughput applications in robotics, healthcare, logistics, autonomous systems, and other fields, meeting diverse scenario demands.


Driven by JetPack 7 at its core, featuring Linux kernel 6.8, Ubuntu 24.04 L OS, and the latest NVIDIA AI software stack, Jetson enables low-latency, deterministic execution of advanced generative AI models, powering physical AI applications. It combines hardware-accelerated computing with system-level optimizations, empowering complex systems like humanoid robots, autonomous machines, and industrial automation systems with responsiveness and intelligent behavior.


With integrated Holoscan Sensor Bridge, MIG support, and a Preemptable Realtime Kernel, among other features, Jetson software significantly enhances the performance and efficiency of tasks like high-speed sensor fusion and motion planning. Supported by Jetson AI Lab and a broad ecosystem, Jetson software drastically reduces time-to-performance for edge AI and robotics applications.


The Jetson Thor platform supports the new Cosmos Reason—an open-source, customizable, 7-billion-parameter reasoning VLM for physical AI and robotics.


Figure 6. Jetson Software Stack



JetPack 7 Designed with SBSA Architecture



With JetPack 7, Jetson software aligns with the Server Base System Architecture (SBSA), bringing Jetson Thor's design in line with industry standards for Arm server designs. SBSA's standardized key hardware and firmware interfaces provide enhanced OS support, easier software portability, and smoother enterprise-level integration. On this foundation, Jetson Thor supports unified installation of CUDA 13.0 across all Arm target platforms, simplifying development, reducing version fragmentation, and ensuring a consistent experience from server-level systems to the Jetson Thor platform.



How Does NVIDIA Isaac End-to-End Accelerate Robotics Development?



NVIDIA Isaac is an open-source robotics platform comprising a series of CUDA-accelerated libraries, frameworks, and AI models for developing various robotic products like autonomous mobile robots, robotic arms, and humanoid robots. Modern robots require an advanced "brain" composed of control, vision, and language models, capable of seamless perception-to-action through real-time processing of multimodal data.


Jetson Thor is designed to run high-load models like Isaac GR00T N1.5, providing humanoid robots with real-time human-robot interaction, spatial awareness, and robust environmental perception. The Isaac platform, working in synergy with Jetson Thor, enables the deployment of scalable multimodal AI at the edge, accelerating robotics innovation in industrial and research fields.

Figure 7. NVIDIA Isaac GR00T End-to-End Accelerates Robotics Development


How to Extract Valuable Information from Edge Cameras with VSS?


The NVIDIA Blueprint for Video Search and Summarization (VSS) in NVIDIA Metropolis provides developers with tools to build and deploy video analytics AI agents. These AI agents can analyze real-time camera streams to enable contextualized real-time alerts, video summarization, and intelligent Q&A functionalities.


VSS supports visual agent applications across multiple domains: visual inspection and worker safety monitoring in manufacturing; enhancing fan interaction and athlete data analysis in live sports broadcasting; and reducing emergency response times for road incident management.



Figure 8. VSS Application Scenarios Include Visual Inspection in Manufacturing, Worker Safety Monitoring, and Reducing Emergency Response Times for Road Incidents



How Is Holoscan on Jetson Thor Used for Real-Time Sensor Processing?



NVIDIA Holoscan is an AI sensor processing platform that provides the full-stack accelerated infrastructure required for software-defined real-time AI. It simplifies the deployment and enhances the scalability of edge AI on enterprise-grade hardware, delivering high-performance edge solutions for real-time AI applications.


Running Holoscan on Jetson Thor allows for secure partitioning and isolation of concurrent AI workflows—a capability that enhances determinism and fault tolerance for mission-critical applications and provides protection against data leakage. This enables developers to advance AI innovation without compromising security, making Holoscan a trusted operational layer for real-time motion control in regulated fields.


Modern robots rely on various sensors for intelligent operation, including cameras, IMUs, and actuators, each critical for the robot's proper functioning. With the NVIDIA Holoscan Sensor Bridge, sensors of any type can be seamlessly connected to the NVIDIA Jetson platform via Ethernet. Jetson Thor supports new Ethernet camera technology that transmits sensor data directly to GPU memory, significantly reducing latency and CPU utilization. This approach replaces the complex logic of traditional drivers with a concise software-defined API, enabling precise synchronization and robust scaling for real-time edge AI applications, whether in robotics, industrial automation systems, or advanced medical systems.



A Comprehensive Ecosystem to Shorten Time-to-Market



The Jetson ecosystem includes over 1,000 partners offering targeted support services to help developers bring solutions to market faster, providing essential assistance throughout the development process.



Table 4. The Jetson Ecosystem Includes Over 1,000 Partners



Figure 9. NVIDIA Partners Offer Flexible Collaboration Models; Developers Can Choose Specific Components and Services Based on Their Design Needs



Get Started Developing Physical AI with NVIDIA Jetson Thor Today



Integrate the Jetson AGX Thor Developer Kit into your existing robot to accelerate software development. Developers can start creating and testing applications without waiting for full system integration.


Join over 2 million developers worldwide in starting your next-generation physical AI project. The Jetson AGX Thor Developer Kit (priced at $3,499) and the production-ready Jetson T5000 module are now available for purchase through NVIDIA's global authorized distributors.


Get started with the NVIDIA Jetson AGX Thor Developer Kit by downloading the latest JetPack 7, and access comprehensive documentation, support resources, and tools through the Jetson Download Center and ecosystem partners.