Festo, an Esslingen, Germany-based automation and industrial controls manufacturer, has helped businesses large and small improve their efficiency by delivering various forms of automation technology to organizations looking to automate difficult tasks. As manufacturing processes become increasingly complex, though, Festo has turned to the power of high-performance computing (HPC) to help better tailor solutions to customers’ individual needs.
“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 Dr. Shahram Eivazi, researcher at Festo and a collaborator on the project. “But with artificial intelligence (AI) 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.”
The Festo team recently started a collaboration with the High-Performance Computing Center Stuttgart (HLRS) through the CATALYST project in order to train robots to perform complex tasks safely. Using the center’s world-class HPC resources and partnering with HLRS staff, the Festo team is developing an AI workflow for training robots based on biological learning principles. CATALYST supports activities aimed at evaluating AI solutions and the eventual convergence of AI and HPC to enable full support of AI workflows on HPC.
When training a machine to “learn” a new behavior, researchers primarily use three different methods. The first two, supervised and unsupervised learning, involve using large amounts of data to train an algorithm to pick out patterns effectively—either specific patterns that a programmer wants them to focus on (supervised learning) or noticing correlations of any kind in a given data set (unsupervised learning).
When training a computer to distinguish between cars and trucks, for example, supervised learning would involve feeding an algorithm many images of both and giving feedback about which are cars and which are trucks. For an unsupervised learning application, though, researchers might show the algorithm many pictures of cars and trucks, but let the algorithm define its own parameters for grouping the images. While it might notice structural differences between cars and trucks and filter the images that way, it might also choose vehicle color as the most important parameter and distinguish all the red vehicles from the blue vehicles.
While these are the most common AI methods, robots being designed to automate complex tasks need to be trained in a more detailed manner. The Festo team uses reinforcement learning to train its algorithm, an approach that draws heavily from methods used in early childhood development. Simply put, researchers train the algorithm by giving it feedback on its decisions. It boils down to sequence of trial-and-error. An algorithm wants to achieve a specific goal, such as getting a robot to tighten a screw, and each time it turns the screw driver in the correct direction, it gets positive feedback, a so-called award (the screw goes deeper into the material), otherwise, there will be negative feedback (the screw falls on the ground, because it is too loose).
Using a mixture of input data from the Festo R&D lab as well as video and sensor data from real-world manufacturing environments, the researchers train the algorithm to replicate behaviors while receiving feedback.
“Once collected, you can take these different data sets and turn it into simulation,” Eivazi said. “We wind up with a large dataset that can show the algorithm what is considered good or bad behavior. 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.”
By tailoring these solutions to specific scenarios, Festo can help clients develop more complex automation workflows that involve humans interacting with robots safely. “We want to make these kinds of interactions safer, so we don’t have to put barriers between robots and workers, because ultimately, we want our systems to support humans,” Eivazi said.
The principles behind using reinforcement learning to train an algorithm sound relatively simple, but the devil is in the details—in order to train its algorithm, the Festo team requires about 70 to 100 terabytes of data (100 terabytes is equivalent to saving roughly 50,000 hours of high-definition video to a computer). Using their own in-house computing resources, the team was unable to efficiently analyze such a massive dataset. By partnering with HLRS, however, Festo researchers can take advantage of the center’s Cray CS-Storm system.
The team knew that it would need GPU accelerators to effectively train its algorithm, and while it had previous experience with accelerators, large-scale simulations require a different approach.
“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?’” To take advantage of that performance, though, the researchers need to build out their software with a larger system in mind.
In order to scale its application appropriately, the team has started closely collaborating with Dennis Hoppe, Head of the HLRS Service Management and Business Processes Division, and his team member, Oleksandr Shcherbakov. The HLRS staffers are working with Festo to port their application to run effectively on HLRS’s resources, and will soon start running their application on HLRS systems.
Having access to raw computational power does not mean all that much if researchers are unable to efficiently move and store these large datasets, though, and with a robust storage infrastructure, HLRS can effectively manage Festo’s data in a secure environment that integrates with multiple computational and data analysis tools.
As the collaboration grows, Festo indicated three main challenges the team will have to overcome. First, the team needs to effectively train its algorithm to get “smarter” as it goes. “Training an algorithm with reinforcement learning doesn’t mean it thinks like a human,” he said. “If I train a machine to pick up something, and move it somewhere else, it learns it. Unfortunately, if you then decide to ask it to cut something or screw in a bolt after the first task, you are basically starting from the beginning again.” He indicated that throughout the project, Festo wanted to investigate ways to reuse datasets for additional training opportunities.
Second, collecting meaningful datasets is a challenge. While simulations can go as fast as processing power allows, conducting experiments in the Festo R&D lab means keeping robots moving at real-world speeds, which, for safety reasons, cannot be too fast for humans to react to or interact with.
Finally, the team has to optimize how to move data containing insights gained on HLRS resources and quickly apply it to real-world manufacturing scenarios. As part of the Deutsches Forschungsnetz, with access to its ultra-high-speed X-WiN network, and through the Gauss Centre for Supercomputing’s (GCS’s) InHPC-DE initiative, HLRS has built out high-speed data transfer infrastructure to German universities, research facilities, and its fellow GCS centers. The center will rely on its experience in building out its data transfer capabilities in order to further improve the data management abilities for countless industrial partners.
“The CATALYST project gives us the opportunity to work closely together with researchers from both academia and industry on real-world solutions that combine AI and HPC,” Hoppe said. “The collaboration with Festo goes beyond applying classical machine learning by focusing on reinforcement learning, which is currently a very active research area. It comes with different hardware and software requirements, making the Festo collaboration an excellent example for evaluating HPC’s capabilities in supporting reinforcement learning.”
Funding for HLRS computing resources was provided by the Baden-Württemberg Ministry for Science, Research and Art, and by the German Federal Ministry for Education and Research through the Gauss Centre for Supercomputing (GCS).