Over the past year, machine learning (ML) and artificial intelligence (AI) have begun dominating conversations around the globe like never before. With ChatGPT (an AI-powered language model) reaching 1 million users in just five days, AI is fast approaching an inflection point that could soon transform every industry. As manufacturers increasingly rely on technological advancements, advanced analytics, and data-driven solutions, AI will help improve efficiency, streamline production, and make operations more cost-effective. More importantly, it will open a new frontier for automating what was un-automatable in the past.
Almost four years ago, I wrote an editorial for Bright comparing the future of manufacturing automation to self-driving cars. While there have been continuous learnings and technological jumps since then, I believe this general concept remains true. By increasing the number of sensors and fusing their output with AI and real-time control–just as is done in the self-driving sector–manufacturing will see a seismic shift in the productivity of currently automatable processes as well as what can be automated. Not only will extremely difficult processes now be possible, but the error rate from unforeseen edge cases and unhandled exceptions will be driven towards zero. This will allow devices that were too complicated or too expensive for automation previously, to begin passing through robotic assembly lines.
Most mature manufacturing processes have low error rates, especially those that are fully automated. This makes sense; the goal of the entire factory is to increase throughput by decreasing the time to make something and increasing the quality. Six Sigma, a famous methodology for quality improvement, is named for the concept of having 3.4 defects out of 1 million opportunities. However, realistically for most consumer goods, there is some statistically significant amount of error, or scrap, that comes from the manufacturing process. An acceptable level is determined by customer and business requirements, and the error, or rework factor becomes part of the cost of the product. Depending on what is being manufactured, this could be anywhere from a very small fraction of one percent to significantly more.
There are products, however, where this would be unacceptable. Think of a jet engine. Why are these assembled by highly trained, highly capable individuals? The loss of a single unit would be catastrophically costly. Frequently, a combination of complexity and cost drive a lack of automation for extremely expensive goods.
Historically, the source of errors in automated manufacturing largely come from edge cases that are not covered in the software; the raw material has changed, and the system does not account for it. A machine is moved after years of service in one location and a vision system now needs a small tweak to its parameters because the lighting is slightly different. While this classical automation can still be a feat of engineering, it is nearly impossible–or at least prohibitively expensive–to catch all the edge cases before it is deployed in the field. AI software, fed by sensor data in real-time, will help change this.
The navigation systems, both vision and controls, will no longer be inextricably linked to the engineer’s experience who programmed them. With small amounts of prompting and feedback, they will be able to quickly explore the space of possible outcomes and edge cases, making intelligent corrections on 1 in a million scenarios and stopping the system before any catastrophic error occurs. The intermediate step, which we are leveraging today, is utilizing AI to detect patterns in processes drifting out of their control limits, without requiring highly skilled manufacturing engineers to sift through a large ammount of data to find these proverbial needles in the haystack.
Various companies and research groups are already using ChatGPT and similar tools to explore how engineers program automated systems today. Frequently, these studies have to do with the “happy path” of the automated system. However, as time goes on, we can think of the relationship between manufacturing systems and AI as that of a convolutional neural network (CNN) and generative adversarial network (GAN)[AM1] [BC2] [AM3] . That is, the AI that drives the manufacturing system will continue to make it better by proactively thinking of, and avoiding edge cases, based on the data it is being fed with. In the future, we aim to limit the current guardrails on these complex systems, and allow them to make corrections on their own.
As AI continues to improve and enable more and more complex behavior, it will become an even more regular part of automation, both in the cloud and on the factory floor. Systems that can understand their surroundings, predict failures, and react in real time will be fused together with hardware to enable more and more complex devices to be assembled by machines. The innovations introduced over the past year will have resounding impacts for decades to come.