DLR: Helping robots to help themselves with Google Cloud

About DLR

The Institute of Robotics and Mechatronics is a part of Deutsche Luft und Raumfahrt (DLR), Germany's national aerospace institute. With more than 150 researchers, it's one of the largest pure robotics research institutes in the world.

Industries: Technology
Location: Germany

DLR's Institute of Robotics and Mechatronics uses Compute Engine instances to train its robots with highly precise simulations and deep learning techniques.

Google Cloud results

  • Trains robots to learn from their environment rather than pre-programming with deep learning algorithms
  • Scales on demand with Compute Engine virtual machines to run complex training simulations
  • Customizes instances to suit particular tasks, with a range of hardware options

Helps to run training simulations 5 to 10 times faster

In May 2018, the Bavarian Minister of Economic Affairs and an invited audience including press and healthcare workers met in the German town of Garmisch-Partenkirchen near the Alps. They gathered round to watch an elderly care worker named Justin pull open a drawer, find a bottle of pills, and bring it to a patient. But why were all eyes on Justin? The answer may surprise you: Justin is a robot. First introduced to the world in 2009, he represents the culmination of nearly a decade's worth of ground-breaking research in both hardware and software. Justin demonstrates touch sensitivity and motor skills to rival a human and the ability to plan and carry out complex tasks without extra programming.

"We wanted to use deep learning techniques to make our robots more intelligent and autonomous. But training these models requires an extreme amount of compute power for training and running simulations. That's where Google Cloud comes into play."

Berthold Bäuml, Head of Autonomous Learning Robots Lab, DLR

Developed by the Institute of Robotics and Mechatronics, part of Germany's national aerospace institute Deutsche Luft und Raumfahrt (DLR), Justin is a humanoid robot, originally designed to operate in hostile environments such as space or Mars. Thanks to real-time 3D modeling capabilities and some of the most advanced arm and hand mechatronics in the world, by 2011 he could throw and catch a ball as well as make a cup of coffee. "We got to the point where hardware wasn't a limiting factor," says Berthold Bäuml, Head of the Autonomous Learning Robots Lab at DLR. "In the last few years, as we began to adapt Justin for terrestrial use, we've really begun to focus on intelligence and learning." To get to that point, the Institute of Robotics and Mechatronics wanted to explore deep learning, a new subset of machine learning that aims to teach AIs how to learn for themselves. But getting the most out of deep learning requires deep resources, so the institute turned to Google Cloud.

"We wanted to use deep learning techniques to make our robots more intelligent and autonomous. But training these models requires an extreme amount of compute power for training and running simulations," says Berthold. "That's where Google Cloud comes into play."

Extreme power, extreme flexibility with Compute Engine

The Institute of Robotics and Mechatronics is aiming to make Justin and other robots autonomous, meaning that they should work out for themselves how to carry out complex tasks in unfamiliar environments, without any extra programming. Berthold and his team believe that robots can learn in much the same way humans do. "When you're born, you start to interact with the environment and you learn how things work," he says. "It's not about programming the system, it's about training it to learn."

In recent years, this has been made possible with new algorithms and techniques such as deep reinforcement learning, where a robot is taught to learn from its interactions with its environment. Unlike humans, Justin and other robots like him can be trained in virtual environments, which simulate the conditions they might experience in the real world. In 2017, when the Autonomous Robots Lab began to use deep learning algorithms, the sheer power needed to generate environments quickly overwhelmed the existing infrastructure. "Our in-house servers weren't enough to handle the kind of things we wanted to do, such as neural architecture searches or physics simulations," says Berthold. "We had to choose between buying new hardware or looking at our options with the cloud."

"We can configure powerful instances very easily with Google Compute Engine, using hundreds of CPU cores if we need to. We have a number of specialized configurations that we can use to spin up instances whenever we need them. We simply couldn't do that with in-house servers."

Berthold Bäuml, Head of Autonomous Learning Robots Lab, DLR

The institute decided on a cloud solution and, after evaluating the leading providers, chose to run virtual machines with Compute Engine because of its combination of affordability, flexibility, and ease of use. Cost was a crucial factor in choosing a solution for a public institution like the institute, and it found that Compute Engine holds its own against the other contenders. Beyond its competitive pricing, Compute Engine also offers a range of hardware options that can be configured with a robust API.

"We can configure powerful instances very easily with Google Compute Engine, using hundreds of CPU cores if we need to," explains Berthold. "We have a number of specialized configurations that we can use to spin up instances whenever we need them. We simply couldn't do that with in-house servers."

Running training simulations and complex tasks on Compute Engine means that the institute can not only add more machines at will, it can also choose different CPUs and GPUs to optimize its configurations. In an on-premises infrastructure, this would be prohibitively expensive, requiring the institute to buy and install new hardware each time it wanted to experiment with a new configuration. On top of that, the simple interface means that researchers, who are familiar with Unix-style environments, are empowered to launch their own instances with ease.

A new kind of learning grounded in Google Cloud

Since moving to Google Cloud, DLR's Institute of Robotics and Mechatronics has made a number of advances with Justin and some of the other robots. For instance, with bigger and more detailed training environments, Justin can now detect and distinguish between materials by touch with more sensitivity than even humans. Another application is the dextrous in-hand manipulation of objects with the sensitive multi-fingered hands where, for example, the robot learns on its own to rotate a cube in its hand without letting it drop.

"With Google Cloud, we can train our robots between five and ten times faster than before and really explore things like deep reinforcement learning," says Berthold. "We can do cutting-edge research, and we're no longer bound by our computing resources. It's been a game changer."

Berthold Bäuml, Head of Autonomous Learning Robots Lab, DLR

Perhaps the most impressive application that the institute is working on is deep-learning based symbolic planning. Robots are given a plan and a set of materials, such as cubes or scaffolding, before working out for themselves how to construct stable structures. This sort of application would be useful in hostile environments such as Mars, where astronauts would need some infrastructure already built if they were to operate there. For Berthold, this kind of ambition has only been achievable with access to the cloud.

"With Google Cloud, we can train our robots between five and ten times faster than before and really explore things like deep reinforcement learning," says Berthold. "We can do cutting-edge research, and we're no longer bound by our computing resources. It's been a game changer."

After last year's demonstration, DLR's Institute of Robotics and Mechatronics has continued to work on integrating its robots in elderly care in Garmisch-Partenkirchen, with trials expected to start in the next year or two. "The elderly people we've talked to are very interested in an assistant who can help with everyday tasks like bringing a glass of water or preparing a simple meal," says Berthold. "The goal is to build a touch-sensitive, highly dextrous, intelligent robot that can safely navigate a human environment."

Meanwhile, the institute has kept a close eye on developments with Google Cloud as new hardware becomes available. It's currently exploring the use of high-performance Tensor Processing Units to save on compute costs but also embed real-time processing capabilities on the robots themselves. Open and clear communication with Google Cloud has been key to the institute's success from the very start. "Our relationship with the Google Cloud team has always been very good," says Berthold. "We get the impression that the team doesn't just see us a customer. Instead, it's genuinely interested in what we can do with Google Cloud technology."

About DLR

The Institute of Robotics and Mechatronics is a part of Deutsche Luft und Raumfahrt (DLR), Germany's national aerospace institute. With more than 150 researchers, it's one of the largest pure robotics research institutes in the world.

Industries: Technology
Location: Germany