AI Will Transform Manufacturing Robotics—Eventually

Many people hear about artificial intelligence and think “robots!” But the great gains in Large Language Models that power ChatGPT, Claude and Gemini are having trivial impact on industrial robots, at least so far. The future looks much better, according to experts in the industry, but not right away.

The LLMs process words and images in ways that enable them to summarize information, ask questions and generate images. The capabilities of the LLMs have blossomed in the past few years, but that does not immediately translate into a factory setting.

“There’s a whole lot of robots out in the world that don’t have any AI component in them whatsoever and are doing meaningful work,” Erik Nieves, CEO of Plus One Robotics told me on a video chat. He said that whatever was “… sitting in your garage was built by robots that don’t know what AI is.”

Welding has been performed by robots in automobile manufacturing since 1967. Key to success there is a controlled environment and large quantity of repetitive operations. A fender weld for a given model car is the same over and over. Car model production runs typically amount to over 100,000 units per year and can be much larger. A customized robotic welder makes sense. But a company making 1000 units of some product may not find a robot cost effective. A robot would make even less sense for field repairs with variable temperature and humidity. That’s true even though robots are cheaper than ever before, according to Nieves.

Advances in machine learning (the mathematical operations behind AI) enable better robotics when there is more data and information. That’s why Nieves’ company specializes in package sorting equipment. There were 161 billion packages shipped in 2022, according to Pitney Bowes. That’s enough data for AI to learn the shape of packages and how to figure out where they are going.

Advances in factory robotics will come from better interconnection for data flows, according to Tyler Bouchard, CEO of Flexxbotics. He told me in a Zoom call that a key problem today is that equipment doesn’t talk to other equipment, nor to a company’s data system. That will improve over time, enabling more data collection. At some point, machine learning will enable the system to “see” the entire production process from raw materials ordering to end product shipping and billing. Things that get connected get smarter was Bouchard’s point, reminiscent of the intelligent computer in Robert A. Heinlein’s novel The Moon Is a Harsh Mistress.

Human talent is another limiting factor in robot intelligence. AI requires smart people knowledgeable in machine learning, which we don’t have enough of today. The large AI companies are hiring all they can find, leaving few left over for less sexy and less remunerative fields such as robotics. Engineers who can work on the mechanical and electrical parts of robots are easier to find, and Nieves reports that the many high school robotics teams inspire people to work in the field after college. A longtime coach of high school robotics teams, Dale Yocum, told me on a phone call that programs are growing, with more students participating every year. Most do not go onto careers in robotics, though they learn a lot about teamwork as well as computer programming, mechanics and electrics.

Robotics will continue to grow in areas with a large number of repetitive operations. The cost of the robots will decline while the cost of human labor will continue to rise (as I’ve argued before, due to demographic change.) That will make robots gradually more cost effective for smaller production runs and in smaller factories.

At some point in the coming years, the machines will communicate with one another and with the company’s data system. The purchasing department will know that a motor will be up for replacement in the next month. The shipping department will learn that 500 units of a product will emerge from assembly and packaging tomorrow. The production manager will see that delivery times for the Model 100 are lengthening, so the system is shifting the Number 18 CNC machine to making shafts for the Model 100. Humans will perform some maintenance and some operations, with managers nodding their heads at the decisions made by the system. Bouchard is optimistic that this day is coming soon. But the AI will probably be ready before the hardware has the appropriate interconnections.