Where is this robotics revolution that we were promised? For decades, countless scientists and science fiction writers have predicted a future in which humans would live side by side with autonomous machines that would understand us and make our lives easier. But the truth is that, despite the increasing number of humanoids capable of speaking and fulfilling the most diverse tasks, the only automaton that, today, seems to have really entered our society is a vacuum cleaner with a shape. Frisbee called Roomba.
And that advances in this field have not stopped since Unimate, the first industrial robot in history, began working on a General Motors assembly line in 1961 . In the last half century, the capabilities of these types of machines and the way they collaborate with humans have been improving. If we still haven’t gotten used to seeing them in our day-to-day lives, it is because, mainly, they are found in factories. In fact, Unimate was an automated arm that weighed almost 2,000 kilos. The automotive giant used it to move large pieces of hot metal and weld them, a task that posed risks for operators.
It was the first time that a machine completely replaced a human at work. The new technologies that emanated from the Industrial Revolution, such as tractors and mowers, made it possible to reduce the number of employees, but did not eliminate them completely, since they were devices that had to be operated or supervised by people. But the Unimate was not a simple machine. He could take on a specific task on his own.
In the last half century, the number of robots that carry out some work in industrial processes has not stopped growing. In 2014, there were already 478 of them for every 10,000 human operators in South Korea, which means that almost 5% of the workforce in that country is made up of these mills. South Korea leads this classification followed by Japan, with 315 automatons, in the same proportion; They are followed by Germany, with 292, according to data from the International Federation of Robotics.
Machines capable of adapting
If there are already so many machines building vehicles, handling electronic components and heavy metals, why doesn’t the same thing happen in stores, in hotels or even in our homes? Why don’t we yet have robotic and android dependents to do the housework? The answer is that, for a machine, it is not easy to manage the unknown .
Most industrial robots tend to face repetitive tasks where, in general, they do not run into situations for which they are not prepared. On assembly lines, the parts they handle are always in the same place, are the same size and are always handled in the same way. To learn how to do it, the robot only needs to know where to act, at what speed to do it, and the specific actions required by the task. Thus, at least, conventional programming works, which the Robotics professor Carlos Balaguer, from the Carlos III University of Madrid, describes as a rudimentary process in which if any external detail changes, you have to start over.
“Robots have the potential to improve people’s lives. They could assist astronauts traveling to Mars, patients in a hospital or workers in factories, “explains Stefanie Tellex, a computer science expert at Brown University, in the US, who is investigating how they could. These devices communicate with us using natural language. Classic scheduling allows them to take on tasks, but not deal with accidents , for example. To do this, they should be able to understand the world and their environment, so that they could react autonomously to any situation.
In the waste treatment industry, there are already specialized machines for the recovery of polyethylene terephthalate (PET) bottles, a plastic used in the manufacture of containers. One of them is Wall-B –its name recalls that of the nice robot from the movie WALL-E–, developed by the Spanish company Sadako Technologies. The ingenuity distinguishes PET containers from all waste by means of an artificial vision system that tries to reproduce what the human brain does when it processes the visual signals that our eyes pick up.
To do this, Wall-B has several cameras and artificial intelligence algorithms that allow it to process the images and understand what it sees. “PET bottles are not always in the same position and tend to arrive flattened and intermixed with other objects,” explains Belén Garnica, head of business development at Sadako Technologies. For this reason, your robot has had to learn to recognize the packages it is looking for and to carry out the necessary movements to retrieve them. Automata like Wall-B end up knowing what to do thanks to the most widely used robotic training technique today: machine learning. This consists of showing you thousands or millions of examples, until, from them, you come to infer a behavior. ” It is about the robot learning from experience without being explicitly programmed for it, and that its knowledge increases with each new task, object or situation it faces, ” explains Balaguer.
This is how the face recognition system works on Facebook, for example. Its algorithm was exposed to millions of photographs that had previously been hand-tagged by people. These had marked in each one the area of the image in which the face was found; the software scrutinized them until it was able to accurately distinguish each individual.
Cyberprofes and student robbers
Thanks to this strategy, robots are gaining autonomy and a certain ability to reason. But if the idea is that they end up making our lives more comfortable, the fact that their learning depends on having a large number of people investing their time in labeling millions of data is not very efficient. To accelerate this process, a team of researchers from Stanford University, in the US, has launched the RoboBrain project – an expression that could be translated as Brain Robo -, with the aim that robots teach each other .
RoboBrain would function as a kind of Google for automatons where each machine would go in search of a specific knowledge previously acquired by another . Thus, you could consult, for example, how to grip a cup correctly by its handle or the best way to manipulate a doorknob to open a door. After downloading that specific knowledge and using it to solve the problem it faces, the robot returns its experience to the RoboBrain database. In this way, the next mill to use the service will have more information available.
“The more data that is provided, the better the performance of the robots,” says computer science expert Ashutosh Saxena, who is responsible for the initiative. In his opinion, whether a robot finds it more or less difficult to learn to do something also depends on the environment. “If you are trained by a programmer for a certain situation, it may not work well if something changes, even slightly,” Saxena explains. And he says: “To get to doing something seemingly as simple as making a French omelette, you would have to take into account dozens of variables, such as the look and layout of the kitchen. And no two are the same ”.
But if a robot learns to manipulate an egg without breaking it, its experience could serve another who needs to change a light bulb, for example. He would still have to figure out a few things, including how to screw it on and off, but at least he would know how to hold it carefully. Saxena believes that this shared knowledge will greatly multiply the capabilities of these machines, which will be seen within ten years. The enormous potential of this strategy has piqued the interest of many experts. The Massachusetts Institute of Technology, through its publication MIT Technology Review , has considered it one of the ten most outstanding emerging technologies of the year.
The aforementioned Stefanie Tellex has launched a similar initiative, the One Million Objects Challenge. The idea is that all the units of a specific type of robot called Baxter that are distributed in different assembly lines all over the world share their skills and form a large database on the handling of very different parts.
One of the companies most interested in robots being able to identify and manipulate things is Amazon. In its huge warehouses, thousands of products of different shapes and sizes are piled up waiting for someone to buy them. And when it happens, an operator must walk the facility to locate and package them. The company, aware of the inefficiency of the process and the loss of time it entails for its employees, organized a contest for robotic warehouse workers. The participating mills had to find, collect and pack 25 different objects, from packets of cookies to rubber ducks.
But none of the 31 candidate robots managed to complete all the tests successfully. For example, paperback books were a big problem for the machine introduced by the University of Alberta (Canada), which was equipped with a three-fingered hand. Because they were so thin, the robot’s camera could barely detect them, and when it finally did, it couldn’t get a good grip on them with its thick limbs.
The meeting, called the Amazon Picking Challenge, showed that no matter how sophisticated a robot’s hardware is, it is useless if the software that controls it does not understand the environment . “There are incredible robots, but they cannot function to the maximum of their capacities, because we still lack the necessary algorithms,” says neuroscientist Demis Hassabis, who leads Google’s artificial intelligence research program. In 2010, Hassabis founded the company DeepMind, specializing in this type of technology, which would later be acquired by the online search giant.
Last March, its AlphaGo system, conceived to play games of go, a millennial strategy game devised in China, defeated South Korean champion Lee Sedol, one of the best players in the world. To train itself, the machine faced itself millions of times. “I didn’t feel like I had the advantage at any time,” Sedol confessed after the match.
It will be years before we live with androids
His story is reminiscent of that of the famous chess player Garry Kasparov. In 1997, this was beaten by IBM’s Deep Blue computer. It was the first time that a computer was imposed on a world champion in the same conditions in which chess tournaments are played. But Google’s achievement is much more complex. A chess player has to deal with an average of 35 possible moves each turn, while in go they amount to 250. “Chess is a game of calculation, but go is too complex; it is largely based on intuition, ”says Hassabis. IBM’s had great computing power, but Google’s machine managed to mimic an intuitive process . It’s a dramatic breakthrough that experts weren’t expecting to happen for a couple of decades.
Obviously, Google’s goal is not for machines to dominate us in board games, just as Amazon does not pretend that a robot knows how to catch rubber ducks. Ultimately, these types of initiatives pursue the development of artificial intelligence capable of understanding the world and dealing with it.
Of course, we are still far from achieving it. It can be seen in self-driving cars, which Google also works on. Despite having been trained to understand traffic signs, sometimes some vehicles do not stop turning in roundabouts because they do not know how to get out of them. Something similar happened to the first versions of the Roomba cyber vacuum when a pet climbed on top and covered their camera. These are just two examples that show the many challenges that robotics still has ahead of it before it faces what professor Carlos Balaguer considers the real great challenge: the development of robot-companions that coexist with us, something that machines still they have to learn .
Images: DHL / Robotcup consortium / Technical University of Berlin
