At Siemens Corporate Technology in Princeton, New Jersey and Berkeley, California, scientists and researchers are working on the future of robotics and automation. Their tools are artificial intelligence and adaptive algorithms. Their goal is to enable robots and machines to learn from experience, and thus to dispense with extensive programming.
Can a robot teach itself to perform a task? That would represent a huge step forward from the current state of the art because until now engineers have had to spend hours laboriously programming each of a robot’s movements into a control system. This hasn’t been a serious problem, because industrial robots usually spend years repeating the same tasks. However, the era of mass production of millions of identical items is drawing to a close because consumers want individualized products that differ from one another in shape, color, and even function.
The automotive industry is a case in point. Some models can have millions of possible equipment combinations. For robotics this spells declining efficiency if each change demands reprogramming.
Learn to solve Problems
In view of this, researchers at Siemens Corporate Technology in Berkeley, California, in collaboration with Professor Pieter Abbeel, at the University of California, Berkeley, are investigating how robots can teach themselves new tasks. “Currently robot programming times are too long to be cost-effective in smaller batch production settings”, says Professor Abbeel.“ In this work we are investigating ways to program a robot just once with the general ability to learn to assemble, and then re-use this learning-to-assemble ability across a wide range of assembly problems.” They are working on the next generation of a promising technology in the intersection of AI and Control, called Deep Reinforcement Learning (DRL), which lets the robot interact with its environment and acquire the necessary skills in a trial-error fashion after multiple iterations. These methods have created significant excitement in the robotic research community and have shown promising results in simulation.
However, their success is much lower when tackling problems in the real world. For example, for a simple task that any three-year-old could do in seconds: insert a metal cylinder into a metal ring, current DRL algorithms can get stuck in suboptimal solutions where the peg is adjacent to the hole, or require way too many robot executions in order to converge. “With DRL, robots can learn by themselves, but they are not given any textbook or material to help them learn.”, said Juan Aparicio Ojea, who heads Siemens’ research group for advanced manufacturing automation located in Berkeley, California.
Robot reads CAD files
CAD design files help here. “CAD files contain information about geometry, final assembly positions, tolerances, etc. It is the perfect book for robots to learn faster”, said Juan Aparicio Ojea. His team’s approach has effectively improved over state of the art robot learning methodologies for tracking the motion plan of the robot, and can solve assembly tasks that require high precision, even without accurate state estimation, in seconds instead of hours. Using this technique, if the part is moved a few centimeters to the side or looks slightly different it takes only a few seconds for a smart robot to figure out that it now has to move its gripper, feeling its way through, somewhat farther to accomplish its task.“
At a high-level, this resembles human behavior. Whenever we encounter a deviation from the expected, we adapt with our sensory-motor control.” says Eugen Solowjow, a Research Scientist in Juan Aparicio Ojea’s team. “Humans have a strong intuitive understanding of the physics of objects around us. We would like to endow robots with a similar form of understanding”, adds Aviv Tamar, a researcher from the University of California.
This work has gained the team the recognition at the prestigious IEEE International Conference on Robotics and Automation 2018, where they were finalists for the best paper in automation, over more than 2000 papers submissions.
AI Expertise spanning across the United States
“We want to assist and augment people, not replace them,” says Juan Aparicio Ojea, who thinks that people should monitor machines instead of competing with them in performing monotonous manual labor. This attitude is shared by many of the scientists and developers at Siemens Corporate Technology in the university towns of Princeton, New Jersey and Berkeley, California. They want to use artificial intelligence to further increase the efficiency of robotic and automation systems, which are among Siemens’ key areas of business.
Two of these scientists are Gustavo Quirós and Arquimedes Canedo. Their goal is to automate automation. More specifically, they want to simplify automation engineering. Today, specialists still spend weeks or even months in front of a monitor writing programs designed to automate new production facilities. Much of this work is repetitive. Moreover, experienced automation engineers have streamlined some of these tasks by using programming and engineering patterns that their younger colleagues still have to learn. Might it be possible to collect this knowledge and feed it into an assistant that would take on simple tasks and help newcomers?
Canedo is convinced that this is a realistic possibility, but that it can’t be achieved in the way that was tried back in the 1980s, when Japanese researchers attempted to systematize all engineering know-how. That effort was not very successful, due, in part, to the engineers’ unwillingness to pass on their knowledge to a computer using consistent models and descriptions of their know-how. Quirós, whose work focuses on the measures that need to be taken during the construction stage of new production facilities, believes that artificial intelligence will enable a new attempt to succeed where the previous one failed. To this end, they are currently developing a cognitive automation engineering system that collects engineers’ know-how and, in a sense, looks over engineers’ shoulders and assists them in new projects. This assistant observes how experts proceed, how they solve problems, and what engineering patterns they use during programming. The assistant learns from the many automation projects it takes part in so that it can provide support in new projects. “Our goal is to increase the productivity, quality, and reliability of automation projects,” says Quirós.
Another researcher Max Wang, who heads Siemens’ research group for automation runtime systems, goes a step further. Wang wants to use artificial intelligence in the runtime to automate decision making in existing automation systems. For example, a train bogie’s sensors might detect suspicious deviations that indicate that a defect is about to occur. An algorithm would classify these deviations, reason about them and the state of the system, draw conclusions from them, and suggest measures on how the operator should proceed. If danger is imminent, it could even automatically and safely shut down the system. According to Wang, the system is currently still learning from sensor data acquired in real operation but processed offline, but future plans include using the system to learn during actual operation.
A Diagnostic Expert in your Pocket
The machine learning-based algorithm developed in Wang’s team is so efficient that it can be used by anyone who has a smartphone. The algorithm that runs in a smartphone app uses smartphone sensors such as microphones and enable the software to determine whether a motor is producing strange noises, for example. This, in turn, would let the software figure out when a specific defect might occur and the possible root-cause. This “expert in your pocket”, as Wang calls it, already exists as a MVP (Minimum Viable Product).
“Increasing the intelligence of machines doesn’t mean that fewer people will be working in tomorrow’s factories,” says Aparicio, who insists that “humans and machines can amplify each other.”