NVIDIA has recently officially launched its next-generation robotics computing platform—Jetson Thor, which is positioned as the “new brain for robots,” specifically designed for physical AI and robotics applications.
1. The Context of Physical AI: Why Jetson Thor Represents a Paradigm Shift
“Physical AI” emphasizes multimodal agents that operate in real physical environments in real-time, rather than just referring to robotic hardware. Compared to traditional cloud-based AI, it requires:
– Low-latency closed-loop: millisecond-level perception-decision-execution cycles;
– High-concurrency inference: running multiple Transformer models for vision, language, control, etc., simultaneously;
– Edge autonomy: stable operation in offline or weak network environments.
Jetson Thor is born for this paradigm, with design goals summarized as: compressing generative AI capabilities that could only run in data centers into a 130 W edge device.
2. Hardware Architecture: Edge Reconstruction of Server-Level Computing Power
| Module | Key Specifications | Significance for Physical AI |
| GPU | Based on Blackwell architecture, 2070 FP4 TFLOPS | FP4 low-precision inference enables Transformers to run in real-time at the edge, with a 7.5x increase in computing power compared to Orin. |
| CPU | 14-core Arm Neoverse V3AE | Designed for real-time control, ensuring motion planning and AI inference run in parallel without blocking. |
| Memory Subsystem | 128 GB LPDDR5X @ 273 GB/s | Large capacity + high bandwidth meet the coexistence of multimodal model weights and sensor stream data; memory capacity is 2x that of Orin. |
| Power Envelope | Configurable 40–130 W | Covers all scenarios from indoor service robots (passive cooling) to outdoor humanoid robots (active cooling). |
| I/O and Functional Safety | PCIe Gen5, USB4, Safety Engine | Supports multiple high-resolution cameras, lidar, and real-time Ethernet, with built-in ASIL-D level safety island. |
3. Software Stack: Making “Large Models” Plug-and-Play on Robots
1. Jetson Platform Services
Pre-integrated containerized microservices support ROS 2, DDS, OPC-UA, achieving zero-friction integration of AI models with traditional robotic middleware.
2. Isaac GR00T N1.5
NVIDIA’s first visual-language-action (VLA) large model, capable of inference directly on Thor, completing the end-to-end process of “human sentence → robot executable action.”
3. TensorRT-LLM & FP4 Quantization
Through FP4 quantization and inference decoding, the same model throughput can be increased by 1.5–2 times, with latency reduced to the 10 ms level.
4. Digital Twin Closed Loop
Synthetic data generated by Isaac Sim → training large models → Thor edge deployment → real data feedback to the cloud for retraining, forming a “data flywheel.”
4. Industry Implementation: Who is the First to “Eat the Crab”
– Humanoid Robots: Agility Robotics’ Digit and Figure 02 have announced the use of Thor as the sole computing node to achieve dynamic balance tasks such as running and climbing stairs.
– Medical Surgery: Medtronic has integrated Thor into laparoscopic robotic arms, using the VLA model to complete “voice command + visual guidance” for intraoperative tissue dissection.
– Agriculture and Construction: John Deere and Caterpillar are evaluating Thor on autonomous tractors and unmanned excavators, aiming to replace traditional x86 + GPU dual-board solutions.
5. Competitive Landscape and Business Strategy
| Dimension | Jetson Thor | Qualcomm RB6 | Intel AMR | Tesla Dojo |
| AI Computing Power | 2070 FP4 TFLOPS | 70 TOPS (INT8) | 150 TOPS (INT8) | Chip-level training, not open to the public |
| Power Consumption | 130 W | 20 W | 40 W | 2 kW |
| Ecological Openness | Full-stack open-source, ROS native | Android/ROS hybrid | Partially open-source | Closed |
| Pricing | $2999 (per thousand units) | About $600 | About $1200 | Not disclosed |
NVIDIA adopts a “high price + ecological lock-in” strategy: seizing the high-end robotics developer mindset through extreme performance, and then forming long-term stickiness through software bindings like Isaac and Omniverse.
6. Conclusion: A New Milestone in Edge AI
The launch of Jetson Thor marks the intersection of two major trends: miniaturization of large models and robotic large modeling. It enables multimodal models at the level of tens of billions of parameters to run in real-time at the edge, upgrading robots from “automated devices” to “physical terminals with general intelligence.” With the evolution of Blackwell Ultra and subsequent Rubin architecture by 2026, we may witness the “ChatGPT moment” in the robotics field within 2-3 years.
“Physical World → Large Model → Physical Action” closed loop
Jetson Thor does not simply allow large models to directly process raw sensor bytes; instead, it uses Holoscan Sensor Bridge + Multi-Instance GPU + Zero-Copy DMA to first “distill” physical signals into large model tokens in GPU memory, before sending them into the Transformer. The entire process is completed within 11 ms and is transparent to developers.
Below, we break down the “ingestion” chain into four levels, along with official metrics and code-level examples.
1. Physical Layer: How Sensors “Plug In”
| Interface | Capability | Typical Sensor Examples |
| 4×25 GbE | 4 channels of 4K60 uncompressed video | 8×Sony IMX-455 industrial cameras |
| PCIe Gen5×8 | 180 MB/s lidar point cloud | 3×Hesai XT-32 |
| SPI-FD | 10 kHz IMU + torque | 6-axis force-controlled wrist + BMI088 IMU |
| USB4/CSI-3 | Audio array, ToF, millimeter wave | 8-mic circular array |
All ports are uniformly protocol-converted by the Holoscan Sensor Bridge chip → directly DMA to GPU memory, with the CPU not involved in the transfer, latency < 1 µs.
2. Data Layer: Turning “Bytes” into “Tensors”
The Holoscan SDK provides two zero-copy paths:
– Video Buffer Protocol (VBP)
Camera RAW → GPU internal Bayer→RGB→Normalization, all completed by CUDA kernel in GPU memory, with no Host-Round-Trip.
– Point Cloud Lidar Pipeline
Lidar UDP packets → GPU kernel for undistortion, voxelization → output `torch.Tensor` directly fed to the VLA model.
Official example code snippet (successfully run on Jetson Thor DevKit):
“`python
from holoscan.core import Application
from holoscan.operators import V4L2VideoCaptureOp, FormatConverterOp
from holoscan.resources import CudaStreamPool
class CameraToTensorApp(Application):
def compose(self):
src = V4L2VideoCaptureOp(self, name=”cam”, device=”/dev/video0″)
fmt = FormatConverterOp(self, name=”convert”, pool=CudaStreamPool(self, 0))
self.add_flow(src, fmt)
“`
Result: 4×4K@60 video → GPU memory waiting for inference, CPU usage < 5%.
3. Model Layer: Multi-Instance GPU Converts “Tensors” to “Tokens”
Jetson Thor brings MIG (Multi-Instance GPU) to the edge for the first time:
It divides the 2560-core Blackwell GPU into 3 instances:
– Instance 1 (1/3 GPU): runs visual encoder (ViT-L 1.1 B parameters) → produces 256 visual tokens;
– Instance 2 (1/3 GPU): runs lidar encoder (Point-BERT 350 M parameters) → produces 128 3D tokens;
– Instance 3 (1/3 GPU): runs fusion Transformer (VLA 30 B parameters) → outputs action tokens.
Hardware isolation ensures that visual encoding does not preempt resources due to control token latency, with end-to-end latency of 11 ms.
4. Execution Layer: Tokens → Physical Actions
Action tokens output by the Transformer (joint angular velocity, end pose) are directly mapped to the real-time domain of the Safety Island through shared memory IPC:
– GPU writes tokens → LPDDR5X fixed physical address;
– Safety Island R52 core reads → EtherCAT master issues servos at 1 kHz;
– Zero-copy throughout, jitter < 50 µs.
5. Developer Experience: Encapsulating “Complexity” into Two Lines of API
“`python
import jetson.thor as thor
# One line to declare sensors
thor.sensor.add(“camera”, type=”gmsl-4k”)
thor.sensor.add(“lidar”, type=”ethernet-hesai”)
# One line to bind large model
thor.model.load(“nvidia/gr00t-n1.5-fp4”)
“`
Holoscan automatically completes:
– Sensor clock synchronization (PTP)
– Data alignment (timestamps, external parameters)
– GPU memory pool pre-allocation
Developers only need to write business logic in Python.
6. Measured Data (Officially Released)
| Metric | Value | Test Conditions |
| Visual → Token Latency | 4.2 ms | ViT-L 1.1 B |
| Lidar → Token Latency | 2.8 ms | Point-BERT 350 M |
| VLA Inference Latency | 3.9 ms | 30 B parameters |
| Control Jitter | 42 µs | EtherCAT 1 kHz |
Conclusion
Jetson Thor uses Holoscan Sensor Bridge to “zero-copy” transfer sensor data into GPU memory, then uses MIG to completely isolate different modal encoding tasks from large model inference, and finally maps tokens directly to the real-time control domain through shared memory IPC.
What developers see is just a clean tensor interface, while the underlying has completed the “Physical World → Large Model → Physical Action” closed loop.
Leading the Industry + Physical AI
Industry Intelligence OfficialAI–CPS
Join the Knowledge SphereIndustry Intelligence Research Institute: Industry OT technology(Automation + Robotics + Craft + Precision Benefits) and new generation IT technology(Cloud Computing + IoT + Blockchain + Big Data + AI)deeply integrated, building a “state awareness – real-time cognition – autonomous decision-making – precise execution – learning enhancement” of Physical AI; realizing industrial transformation and upgrading, driving business value innovation and creation of interconnected industrial ecosystems.
Physical AI as the core driving force of the fourth industrial revolution, will further unleash the enormous potential accumulated by previous technological revolutions and industrial changes, creating new powerful engines; reconstructing design, production, logistics, services, and other economic activities, forming intelligent demands across various fields from macro to micro perspectives, giving rise to new technologies, new products, new industries, new business models; triggering significant changes in economic structures, profoundly altering human production and lifestyle, and achieving a comprehensive leap in social productivity.
Today, in the era of industrial intelligence technology, practitioners must understand how to fully integrate Physical AI into the entire company, products, and business scenarios, leveraging Physical AI to achieve digitization, networking, and intelligence, realizing a new layout for industries, a new construction for enterprises, and a renewed vitality.