At the recent Advantech Edge AI Conference in Taipei, attendees heard about plenty of positive trends for the future. Miller Chang, President of Embedded Sector, Advantech, noted that the Global EdgeAI market is estimated to grow to an incredible $196.6 billion for 2034. Hardware will lead there with AI accelerators, and software and services will start to grow faster as complexity rises. The challenge, Chang said, is that limited standardization increases integration complexity, security complexity, and deployment costs.
Within industry, Chang said that Cloud AI wins with centralized model training and model performance/capability. Meanwhile, Edge AI excels with low-latency execution, on-site local intelligence, and real-time decision making. And Physical AI, the latest buzzword across the industrial sectors, incorporates perception, reasoning, and action.
Getting physical AI up to speed on the factory floor
The conference saw myriad high-level speakers, including Deepu Talla, Vice President, Robotics and Edge Computing, NVIDIA, who noted that while automation has existed for more than 60 years, we still haven’t figured cracked the code for bringing autonomous capability into physical AI and robotics.
“You can test it in many different ways, you can ask how many robots, industrial robots, robots, AMRs, with any sort of autonomous capability, how many are shipping right now annually? You could pick a number — one million? But then you ask yourself, ‘What is really the opportunity that is in front of us?’ The answer is that it’s billions, if not tens of billions. So, we are at least a factor of 1,000 if not 10,000 times away from making this happen,” he told the audience.
Miller Chang, President of Embedded Sector, Advantech and Deepu Talla, Vice President, Robotics and Edge Computing, NVIDIA.
Talla also noted that there are two problems we need to solve in order to make physical AI work. First is accuracy. While accuracy has gotten better and better over the last three years or so, there is still always a human in the loop, because we still haven’t cracked the accuracy code. The second is easy integration. Talla said that once the robots become reasonably intelligent, they still have to integrate into the real world. There are many existing workflows already present in the average manufacturing plant. Think of the various things, from protocols to PLCs to optical inspections to factory operating systems, that have been created over the last 40 years. These robot brains will need to play nice and integrate with all of these systems.
“We are on a quest to build a ChatGPT-type or reasonably good enough general-purpose model that is accurate enough for robotics … in the case of robotics, it’s not about training the model, because — where do you even get the data to train the model? You can’t take YouTube videos and make a good enough industrial robot. You probably have a good dancing robot, but it’s not going to be really solving precise tasks for the accuracy that you need. So, how do you solve this problem? Well, you can go and collect lots of real-world data that’s happening right now,” said Talla. “There are people collecting from heavy operation to implementation learning to seeing general-purpose videos. All of those are necessary, but actually are not sufficient, and it turns out the answer is again using AI to take whatever real-world data is available to create many scenarios, and that’s where we spend a lot of time right now.”
NVIDIA’s goal is for robots to be able to work seamlessly in a variety of industrial settings, leveraging technologies like Cosmos and Unity to create these generalist models.
And here in the U.S., robots are back to double digit growth, so solving these challenges is of great importance. According to statistics just released by the International Federation of Robotics (IFR), the number of industrial robot installations in the United States rose by 11% year-on-year, to reach 38,000 units in 2025. IFR said that this is being driven by robust growth in the food industry and other non-manufacturing sectors — although the automotive industry remains the largest adopter and reached 13,500 units.
“The United States are back on the growth track,” said Takayuki Ito, President of IFR. “While automotive achieved its third-best result in seven years, the data highlights a growing demand for flexible automation in the food industry: Adoption in this sector surged by 30%, now ranking alongside metal and machinery and electrical-electronics, all with approximately 3,000 installations in 2025.”
Playing well together
Advantech’s industrial AI software platform is called WEDA (WISE-Edge Developer Architecture) and impressively is able to integrate with NVIDIA NemoClaw (as well as other architectures on diverse computing platforms from Qualcomm, Intel, and AMD) to develop autonomous operations applications for manufacturing and other industries. The company said, “WEDA accelerates AI development and deployment at the edge through a unified development architecture across chips and operating systems, integrating APIs, containerized deployment, digital twin simulation, and AI lifecycle management.”
At the COMPUTEX show, held the same week, Advantech showcased how AI is moving into these physical environments. The company’s latest Edge AI portfolio includes Edge AI software and robotics platforms; industrial automation and intelligent systems; and iHealthcare, iRetail, and hospitality applications. These represent the company’s strategy across computing platforms, ecosystem collaboration, and vertical industry applications.
K.C. Liu, Chairman of Advantech.
“The next stage of AI is not only about breakthroughs in computing power and models, but also about whether AI can truly enter industrial environments and become a new infrastructure for enterprise transformation” said K.C. Liu, Chairman of Advantech. “With its long-standing commitment to edge computing and its position at the forefront of real-world AI implementation, Advantech will continue to leverage Edge AI, AI Agents, and WISE solutions to connect the global partner ecosystem and accelerate AI’s expansion from the cloud to the edge, and from the digital world into physical environments.”
Advantech
advantech.com/en-us