For nearly a hundred years, German-based firm Festo has delivered industrial controls and automation tools to its clients, growing to over 20,000 employees and billions in revenue. Over the last century, though, the automation demanded by those clients has grown increasingly complex – and now, Festo is working with the High-Performance Computing Center Stuttgart (HLRS) to use HPC to train safer robots for automation.
“Festo has years of experience with automation, and until recently, these processes were more or less built once in a facility, then machines perform the tasks they need to do,” said Shahram Eivazi, a researcher at Festo, in an interview with HLRS’ Eric Gedenk. “But with artificial intelligence and other new tech, people are starting to ask for more custom-made solutions in their factories. Automation processes we have developed might need to be changed and tweaked for a company’s specific needs, and that means that these systems have to be adaptive so they can change in a reasonable amount of time, while also being safe and interactive with humans that are involved in the manufacturing process.”
Using data from Festo’s labs as well as observational data from manufacturing facilities, Festo’s researchers train a reinforcement learning algorithm. “We wind up with a large dataset that can show the algorithm what is considered good or bad behavior,” Eivazi said. “Using this method, we can achieve roughly 80 percent of the performance we want without actually ever touching a real environment. Then the last 20 percent of the work is tuning it to a specific environment for a specific need.”
That last 20 percent is tricky – and key to making the resulting automation mechanisms safe for human interaction, where necessary. “We want to make these kinds of interactions safer, so we don’t have to put barriers between robots and workers,” Eivazi said, “because ultimately, we want our systems to support humans.”
The overall computational burden of training those algorithms with reinforcement learning is high, with each training session applying up to 100TB of data. At Festo, their in-house HPC resources weren’t up to the task. “We had experience with GPUs, but always in small clusters—3 GPUs and 100 CPU cores,” Eivazi said. “As researchers in industry, we have limited access to large-scale computational resources and we already passed the point of training what we can on in-house resources, and coming to work with HLRS lets us answer the question, ‘What if we have access to thousands of CPUs instead?’”
Now, Festo’s researchers are working with HLRS to port their algorithms to HLRS’ much more considerable resources, which include the 19.3 Linpack petaflops Hawk system. This work is ongoing, and the team aims to improve the process not only through superior computation, but through optimization of the training algorithms and data collection processes.
“The collaboration with Festo goes beyond applying classical machine learning by focusing on reinforcement learning, which is currently a very active research area,” said Dennis Hoppe, head of the HLRS Service Management and Business Processes Division. “It comes with different hardware and software requirements, making the Festo collaboration an excellent example for evaluating HPC’s capabilities in supporting reinforcement learning.”
To learn more, read the reporting from HLRS’ Eric Gedenk here.
Header image: automation experiments in the Festo R&D lab. Image courtesy of Festo.