Compared to humans, robots are slow learners. But thanks to research now being conducted collaboratively between Siemens Corporate Technology and the University of California at Berkeley, the ability of robots to pick up new skills could expand rapidly. Take cloud robotics, for instance, which potentially gives robots the power of an army of servers. Instead of being limited to local computing resources, cloud computing opens the door to using computationally-intensive resources such as deep learning. Other advantages include the potential to leverage the benefits of open source software, parallelize computations, and, most exciting, share knowledge with other automation systems while keeping data up-to-date.
The Future of Manufacturing — Inside Siemens’ Labs
A Cloud for Robots
Robots are on track to becoming speed learners. Working with the University of California, Berkeley, Siemens researchers are focusing on how robots can harness the power of the cloud to learn – and teach themselves – specialized tasks.
“Today, robots just have to pick and place the same object over and over,” says Juan L. Aparicio, who heads a Siemens research team in advanced manufacturing automation, located in Berkeley - California. “But in the near future, as we move toward lot-size-one, it will be unworkable to reprogram all of a robot’s movements for each new object. And in order to figure out how to grasp new objects on their own, robots will need the cloud.”
Knowledge Sharing among Systems
In view of this, Aparicio and his team are focusing on how robots can harness the power of the cloud to learn – and teach themselves – how to grasp unfamiliar objects. “Grasping is one of the most computationally-intensive activities a robot performs,” explains Aparicio. “But if you give a robot the autonomy to find the right movements from the cloud, it will figure out how to perform the task more quickly and efficiently.”
Working with UC Berkeley’s Dexterity Network (DexNet), a cloud-based grasping database, the Siemens team is providing specialized knowledge regarding manufacturing, as well as methodologies for encapsulating vital IP when sharing cloud-based information. “We are now integrating vision and are generating a synthetic database that includes 6.5 million images combined with 3D models – all of this with a focus on grasping,” says Ken Goldberg, the UC Berkeley professor behind the DexNet platform.
The Siemens team, which focuses on bridging the gap between academia and industry, is also involved in a related project with UC Berkeley’s Professor Pieter Abbeel. The project focuses on deep reinforcement learning for robots. “Deep reinforcement has had great success in games and simulated environments, in this project we are figuring out how to bring the same success to learning by real robots,” says Abbeel. Adds Aparicio: “This project is one of the few examples where deep nets are being applied to control of physical systems. We are designing controllers to enable a robot to learn novel tasks with minimal programming. The next step will be to integrate both efforts, leverage the cloud for knowledge sharing between systems, and give robots the ability to learn and generalize knowledge locally.”