In today’s manufacturing plants, the division of labor between humans and robots is quite clear: gigantic, automated robots are typically confined in metal cages, operating massive machinery and performing repetitive tasks, while humans work in less hazardous areas to perform work that requires more detail.
But according to Julie Shah, a Boeing career development assistant professor of aeronautics and astronautics at MIT, the factory floor of the future could feature humans and robots working side by side, helping each other with shared tasks. Shah predicts that robotic assistants will perform tasks that would otherwise complicate human work, especially in aircraft manufacturing.
“If a robot can provide tools and materials so that a person doesn’t have to walk around the part and get back to the plane, you can significantly reduce that person’s idle time,” says Shah, who leads the Interactive Robotics Group at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). . “It’s really hard to get robots to do the precise finishing tasks that humans do really well. But providing robotic assistants to do non-value-added work can actually increase the productivity of the entire factory.”
A robot working in isolation must simply follow a set of programmed instructions to perform a repetitive task. But working with humans is a different matter: Each mechanic working in the same position at an aircraft factory, for example, may prefer to work differently – and Shah argues that a robot assistant would have to effortlessly adapt to a person’s particular style to be versatile. practical utilize.
Now Shah and her colleagues at MIT have developed an algorithm that allows a robot to quickly learn a person’s preferences for a specific task and adapt accordingly to lend a hand complete it. The group uses the algorithm in simulations to teach robots and humans to cooperate, and will present their findings at the Robotics: Science and Systems Conference in Sydney in July.
“It’s an interesting machine learning and human factors problem,” says Shah. “With this algorithm, we can significantly improve the robot’s understanding of a person’s next likely actions.”
Lifting your wings
As a test case, Shah’s team looked at spar assembly, the process of building the main structural element of an aircraft wing. In a typical manufacturing process, the two parts of the wing are matched. Once in place, the mechanic applies sealant to the pre-drilled holes, drives screws into the holes to secure the two pieces, and then wipes off the excess sealant. The entire process can be highly customized: for example, one mechanic may choose to apply sealant to each hole before driving a screw, while another may want to completely finish one hole before moving on to the next. The only limitation is the sealer, which dries in three minutes.
Scientists say robots like FRIDA, designed by Swiss company ABB, can be programmed to assist in the girder assembly process. FRIDA is a versatile robot with two arms that can perform a wide range of motion. According to Shah, it can be manipulated to drive screws or paint the holes with sealant, depending on a person’s preference.
To enable such a robot to predict human actions, the group first developed a computational model in the form of a decision tree. Each branch on the tree represents a choice the mechanic can make – for example, after applying sealant, should he continue driving the bolt or apply sealant to the next hole?
“If a robot inserts a screw, how confident is it that the person will then insert the screw, or will they just wait for the robot to insert the next screw?” Shah says. “There are many branches.”
Using the model, the group conducted experiments on humans, training a laboratory robot to observe an individual’s chain of preferences. Once the robot learned a person’s preferred task sequence, it quickly adapted by applying sealant or attaching a screw according to the person’s particular work style.
Working side by side
Shah says he envisions robots and humans in real-world manufacturing settings, receiving an initial training session off the factory floor. Once the robot learns a person’s work habits, its factory counterpart can be programmed to recognize the same person and initiate an appropriate task plan. Shah adds that many workers in existing factories wear radio frequency identification (RFID) tags, which is a potential way for robots to identify people.
Steve Derby, associate professor and co-director of the Center for Adaptable Manufacturing at Rensselaer Polytechnic Institute, says the adaptive algorithm developed by the group moves the field of robotics one step closer to true collaboration between humans and robots.
“The evolution of the robot itself has been far too slow in every respect, whether in terms of mechanical design, controls or programming interface,” says Derby. “I think this paper is important — it taps into the whole spectrum of things that need to happen for humans and robots to work side by side.”
Shah says robotic assistants can also be programmed to lend a hand in medical facilities. For example, a robot can be trained to monitor lengthy procedures in the operating room and anticipate the surgeon’s needs by administering scalpels and gauze, depending on the doctor’s preferences. While such a scenario may take many years, robots and humans can eventually work side by side using the right algorithms.
“We have the hardware, the sensors, and we can manipulate and see, but unless the robot really develops an almost fluid understanding of how it can help a person, that person will just get frustrated and say, ‘Never mind, I’ll just do it.’ go pick up the piece yourself,” says Shah.
This research was supported in part by Boeing Research and Technology and conducted in cooperation with ABB.