Physical AI News (Japan)

Physical AI, robotics and compute — reported from Japan.

日本語
· Research & technology· 日本語で読む

Seeed Wiki Publishes Guide to Building a Local AI Agent That Runs on Jetson Orin

Seeed Wiki introduced a local browser agent that runs on Jetson Orin. Laya selects the next action, while Qwen3.5-4B parses requests and generates text. In a Trip.com flight-search demo, results appeared in 32.8 seconds.

What happened

On November 4, 2026, Seeed Wiki published instructions for building a local browser agent that runs on NVIDIA Jetson Orin. In a demonstration, a Jetson Orin NX 16GB installed in a reComputer Industrial J4012 was used to search Trip.com for the cheapest nonstop flights from Shenzhen to Shanghai. It took 32.8 seconds to display the options.

Why it matters

The demo, which used Jetson Orin NX 16GB to operate a browser and display flight-search results from Trip.com, provides a concrete example of a local agent setup on an edge device. It also outlined the models’ respective roles and the conditions for replacing them, but did not demonstrate completing a booking or operating on other websites.

Analysis

A key feature of the demo is that browser operation is divided between a decision model that selects the next action and a model that parses requests and generates text, rather than being handled by a single large model. The decision model can also be replaced, provided it returns operation instructions in JSON format. The setup therefore demonstrates not only agent execution on Jetson but also an architecture in which the model responsible for action selection can be swapped. However, the ability to replace the model does not mean it will perform with the same accuracy across all websites and tasks. The demonstration covered flight searches and whiteboard operations on Trip.com. It did not show a completed purchase or booking, and the flight search was limited to nonstop flights from Shenzhen to Shanghai. The 32.8-second completion time was measured in this demonstration and does not indicate performance on other devices, networks or websites. The local-execution setup was demonstrated, but the materials do not establish whether it works in an environment with no network connection or how it handles changes to a page or login requests. In Marubun to sell RealMan robot arms and humanoids in Japan, we reported that remote-operation and AI data-collection solutions were part of the product lineup. This browser agent performs operations on a screen; it does not provide the same functions as remote robot operation or the collection of real-world work data. Considered together, the two examples show that systems for translating natural language into actions face different environments and validation challenges in web interfaces and robot operations. The materials do not establish whether they can be treated as a shared platform. Areas to watch include search success rates, recovery from operational errors, the range of supported websites, variation in processing time and performance when models are replaced. Safety questions include which operations the agent is allowed to perform and how actions such as purchases or entering personal information are restricted. The article’s demonstration provides a concrete example of the system in action, but the materials do not include evaluation results on these points.

Sources

Translated from the Japanese original: /news/seeed-wiki-jetson-orin-uede-ugoku-rookaru-ai-eejiento-no-kou-79b5ea. The Japanese version is authoritative if they differ.