Robots that learn on the job

From the October 2023 print edition

Generative AI, popularized virtually overnight through the introduction of ChatGPT, is familiar to many as a web-based tool that gathers and formats information for humans.

Lesser known is that in industry, the technology also serves as a tool for enabling skilled workers to transfer their expertise to robots and other mechanical devices.

“One of the most useful applications of AI in industry is that it allows a robot to learn from humans,” says Dr. Alexander Wong, University of Waterloo engineering professor, Canada Research Chair in the area of artificial intelligence, and a founding member of the Waterloo Artificial Intelligence Institute.

The challenge with traditional programmed robots has been that they lack the intuitive flexibility of a human worker. This is getting increasingly critical as manufacturers and supply chain operators adapt to a high-mix, low-volume marketplace where changes in the work are far more frequent.

“Traditionally, all these tasks were pre-programmed, but that doesn’t work well when there’s variation,” says Wong. “If there’s an edge case or an outlier, it doesn’t adapt very well.”

A related problem is that programmers don’t typically understand the tasks they are programming. “When you have developers programming tasks for the shop floor, they don’t have the expert knowledge of the domain experts who actually do that job,” says Wong. “What AI does is allow for a human who understands the task to guide and show the robot how to do
it in a human-intuitive way.”

For example, an operator can guide a robotic arm to pick up an object, and the robot’s AI “brain” can learn from the different trajectories that an experienced person takes it through. “This is essentially a no-code approach to programming,” says Wong.

Robotics manufacturer OnRobot, which specializes in manufacturing a wide assortment of tools and software for collaborative applications, is focusing on improving the interface between machine and skilled worker. “Our mission is to bring automation closer to the end user,” says Kristian Hulgard, general manager – Americas at OnRobot, based
in Irving, Texas, “and we’re actively using AI to bridge that gap.”

This alleviates manufacturing’s perennial skills shortage on two levels. “We have less and less people who want to work in manufacturing,” says Hulgard, “so we have a growing demand
for automation. But we don’t have a growing supply of robotic engineers. So we have to be able
to empower the end user or operator to deploy robots without some massive education, and that’s where AI becomes very useful.”

When programmers are needed, AI also speeds up their work. “Not only does AI make programming understandable from an operator’s perspective,” says Hulgard, “but it also handles the difficult parts of robot programming – signal exchange, handshakes with machines, communications protocols – all these things that require expertise. An AI can do things faster than a human programmer.”

Reduced training load
While AI effectively hands the bulk of robotic programming over to the operators on the shop floor, it can leave these operators with a voluminous teaching load. “One of the key limiters when users are training robots is that there are a lot of different examples that the robot has to be trained on,” says Wong.

“Even if the end user is just guiding the robotic arm, covering all the different scenarios would require a lot of human labour. So that’s where generative AI comes in.”

Wong’s research group at University of Waterloo is collaborating with colleagues at Karlsruhe Institute of Technology (KIT) in Germany in a project, named FLAIROP, to apply generative AI to manufacturing automation environments.

“The question we asked is, ‘what if we actually generate these scenarios in a simulated environment so that an AI can learn from it?’” says Wong.

The project has taken on one of the most variable tasks – bin picking. Items in bins can present themselves to a robotic camera in millions of possible ways, making it impractical to train the robot manually for each scenario. The project is experimenting with a digital twin of a bin environment, which serves as a virtual classroom for the robotic “brain” to learn from an unlimited number of combinations.

“I can take random objects that we have in a manufacturing facility and randomly place them as a bin as an example, and then instruct the robotic arm to learn from it,” says Wong. “Because it’s all simulated, I know exactly where the arm should grasp. And I teach the AI, ‘hey, when you see this scenario, here’s what you should do.’”

This makes it possible to undertake robotic training tasks that would be far too large for
a human trainer to undertake. “Imagine a scenario where an AI can just keep generating data and learning from it,” says Wong, “and imagine an AI training itself for millions upon millions of years to be better. Of course, millions of years in the virtual world is maybe a week or two, or maybe a month.”

Moving forward with AI
The use of AI to improve user interfaces and reduce the involvement of programmers is also taking place in other aspects of automation. “One of the key challenges in production is that the type of data being generated can change frequently, whether it’s due to new sensors that are being installed, renaming of sensors, calibration of equipment, moving around of equipment, et cetera,” says Greta Cutulenco, CEO of Acerta AI, which provides AI-driven predictive quality solutions. “To address this, we have been automating how our platform can detect and adjust to these changes and simplifying the process for our users to quickly remap the data without having Acerta’s engineers get involved.”

Hulgard sees a general trend where more and more industrial products will be software enabled. “I think we can expect that automation manufacturers will make it easier to use automation,” says Hulgard. “Every new product will have software, not just a valve, and it’ll have some sort of application programming interface (API) to make it easier for the end user to install and use. And AI is a normal part of control software now.”

Electric vehicle (EV) production facilities are likely to be a hotbed for this new technology. “Many new EV production facilities are being built and ramped up in Canada,” says Cutulenco. “This is a great opportunity to put digital capabilities into new lines from the start, and a perfect time to start leveraging predictive quality technology. Plants face a lot of challenges during ramp up time, and there is limited experience still with the new EV parts being produced – this is a great opportunity to leverage advanced analytics and predictive data driven technologies to help ensure launch success.”