More and more drones, service robots, and visual inspection equipment are no longer just running simple detection algorithms. Instead, they need to run large language models and multimodal visual models locally to complete environmental understanding, semantic analysis, and real-time decision-making. In the past, such capabilities were mostly dependent on more powerful high-end hardware. However, NVIDIA's newly launched Jetson Orin Nano™ 2 has brought cutting-edge generative AI capabilities down to entry-level embedded robot platforms.

Edge-end physical AI is rapidly evolving. More and more drones, service robots, and visual inspection equipment are no longer just running simple detection algorithms, but need to run large language models and multimodal visual models locally to complete environmental understanding, semantic analysis, and real-time decision-making. In the past, such capabilities mostly relied on more powerful high-end hardware. However, NVIDIA's newly launched Jetson Orin Nano™ 2 has brought cutting-edge generative AI capabilities down to entry-level embedded robot platforms.

Why is a new generation of entry-level edge computing module needed?

There is an interesting change happening in AI models: lightweight model technologies are constantly maturing, and the accuracy of small-sized cutting-edge models has caught up with those of the past large models. This means that many multimodal models that could only run on cloud or high-performance computing boards now have the opportunity to be deployed in devices with limited volume and power supply constraints.

But to truly run models on physical devices, it's not just about having sufficient computing power. Smart drones, household robots, and industrial visual terminals have strict hardware requirements: they need to be small in size, have controllable power consumption, and ensure low inference latency to achieve real-time perception and decision-making. This is precisely the positioning of Jetson Orin Nano 2: an entry-level robot computer for millions of developers, balancing computing power, power consumption, and size.

NVIDIA's Robotics and Edge AI Vice President Deepu Tallar mentioned: "Lightweight cutting-edge models are unlocking real-time intelligence for edge devices. Jetson Orin Nano 2 is designed to enable millions of developers worldwide to obtain the performance and energy efficiency needed for real-time inference in robots and intelligent vision devices." Companies such as Cognex, Doosan Infracore, Matic, and Wing have already begun to evaluate and adapt to this module.

Hardware specifications: doubled computing power, lower power consumption

Jetson Orin Nano 2 is equipped with an 8-core Arm CPU and 8GB of memory. Its AI computing power can reach up to 78 TOPS (trillion operations per second). With a high-performance and energy-efficient hardware form factor, it significantly enhances AI computing and video processing capabilities.

The previous generation of Jetson Orin Nano Super was launched in December 2024, and the development kit consists of a Jetson Orin Nano 8GB system-on-module (SoM) and a reference carrier board. The SoM is equipped with a NVIDIA Ampere architecture GPU with a tensor core and a 6-core Arm CPU, capable of implementing multiple concurrent AI application pipelines and high-performance inference. It can support up to four cameras, providing higher resolution and frame rate than previous versions.

Compared to the previous generation product, Nano 2 has made two key upgrades: improved tensor core + higher memory bandwidth, ultimately bringing two core benefits:

1. Inference performance is directly doubled, and the compact form factor remains the same, allowing the old carrier board to be reused;

2. In 15W power consumption mode, achieving the same performance while reducing power consumption by 40% compared to the previous generation.

For battery-powered devices (drones, mobile robots), the value of reduced power consumption is even more important than a simple increase in computing power. Lower power consumption means longer battery life, less heat dissipation pressure, and reduced design difficulty for the entire system.

At the software level, it fully complies with the NVIDIA Jetson open-source software stack and the Agent intelligent agent skill suite. Developers can directly deploy memory-optimized edge large models, such as Cosmos, Nemotron, Gemma 4, Qwen3, etc., which are open-source large language/visual language models, and run on the board, quickly building multimodal edge AI applications.

The Jetson Orin Nano 2 module and developer kit are expected to be officially launched in the first half of 2027.

Real-world application cases: from delivery drones to household cleaning robots

The value of hardware ultimately depends on whether it can work in actual scenarios. Many enterprises are conducting product verification based on Orin Nano 2.

Wing (a drone delivery company under Alphabet)

Wing's delivery drones have previously been using Jetson Orin Nano Super. They plan to migrate and evaluate Orin Nano 2, aiming to enhance the real-time perception and inference capabilities of the onboard end. Drones need to quickly identify obstacles, pedestrians, and ground environments in the air, and the speed and power consumption of local AI inference directly determine the safety and endurance of the delivery. With the new module, it is hoped to achieve a faster response and more energy-efficient delivery drones.

Matic Robots household robots

The household environment is a highly dynamic changing scenario: furniture movement, scattered items, people walking. The robot cannot only do simple obstacle avoidance. Matic hopes to run large models locally on the cleaning robot to achieve dialogue AI, gesture recognition, semantic mapping, understand the layout of the family environment, and complete intelligent autonomous cleaning. The small form factor of Orin Nano 2 and the edge multi-modal inference capability exactly match the hardware constraints of household robots.

In addition to end manufacturers, a large number of domestic and overseas ecosystem manufacturers (Yongyang, Linghua, Advantech, SiShu Technology, etc.) have already started developing supporting carrier boards, complete machines, customized software and reference solutions to facilitate developers to quickly select, reduce the amount of underlying hardware development work, and accelerate product implementation.

Conclusion

For edge AI developers, in the past, when using Jetson, they mostly ran traditional CV detection and recognition; but with this generation, models with smaller parameter sizes such as LLMs and visual language models can truly run on the board in real-time.

That is to say, the release of Orin Nano 2 signals that multimodal large models are officially moving towards entry-level mobile embedded hardware. Coupled with doubled inference performance, 40% reduction in power consumption at the same performance level, and compatibility with the original form, for mobile robots, drones, and embedded vision developers, this will be a new hardware option worth paying attention to.

Disclaimer: This article is from the internet. If there is any dispute, please contact customer service.