ZEN Robotics

Hi, Timo! What’s the story behind you and robotics?
I joined Zen Robotics two years ago. I have twenty years’ experience in R&D (Research and Development) in different kinds of roles. I came to Zen Robotics from General Electric, where I had been running the R&D side, managing about two hundred and fifty people. Then I decided that the robotics business was a very interesting field, as well as artificial intelligence, so that’s why I joined this company. There’s great business potential here, so that makes it a big and interesting challenge for me. The ability to grow got me into this business.

Is your background in engineering or in AI (Artificial Intelligence)?
My background is in different kinds of engineering, I’m more focused on how to run R&D (Research and Development). I myself am not an AI developer, but of course, after two years in the company, I have learnt many things: both about artificial intelligence and our products.

Of all the possible uses for robots and AI, why was trash-sorting selected as a focus?
We had three company founders with backgrounds in artificial intelligence and robotics. Some of them with an entrepreneurial background and the others with a marketing background. They wanted to find an application for intelligent robots, so they began by contacting many of the biggest Finnish companies and asking them what kind of problems could be solved for them through smart robotics. In 2009, about two years after they founded this company, they came up with the idea of applying robotics to waste-sorting.

The sorting of construction and demolition waste is a very good application area because it has complex enough problems, but it’s still something feasible to accomplish with AI. It’s a way for us to show that we can make a smart robot that works in unstructured environments. In robotics, it’s usually much easier to make a robot that can work in the lab: it’s a structured environment and you know everything in advance. So taking the next step and making it into an actual product is something that we really wanted to do to really utilize the technology and skills that we have.

How efficient is robotic sorting when compared to traditional methods of sorting?
With traditional sorting, humans stand next to conveyor belts and manually pick out the objects. But robots are much stronger than humans, and they are also faster. But the point, really, is that standing next to a conveyor belt picking things is not the nicest kind of work and it’s difficult to find people to do it, so in that sense, robots fill a need: they can take over those kinds of tasks. Because it’s unpleasant, in most countries, not all the waste is sorted, and some of the waste is just sent to a landfill or incinerated. There aren’t any really good competing methods for sorting. There are similar methods, but they have their own problems in comparison to robotic sorting.

What kind of waste is sorted by the robots and what’s the process like?
The waste is mainly construction and demolition waste. There is also some industrial waste, leftover furniture when offices are emptied, for example.

The waste is brought in by trucks into the sorting facility and unloaded onto the floor. An excavator picks up the waste and drops it onto the first conveyor belt. We have a ballistic screen that does the pre-processing of the waste (a process run by another company’s machines) which divides it into three groups: two dimensional materials (like foils), objects that are too small, and objects that are good for robotic sorting. Objects which are too small fall down into waste piles, and the leftover objects that are good for robotic sorting come onto another conveyor belt towards the robots.

How does the robot work out what’s what?
There is a sensor box, and inside the sensor box, we have different cameras. We have cameras that see the shape of the object; laser, line, and 3D cameras. Then we have normal visual cameras and also spectral cameras that can see the material of the object. So based on the spectral wavelength that comes back from the object, different materials are identified. What we do with the sensor boxes is to combine the data from the different sensors and feed that into an artificial intelligence algorithm that determines what the material is, whether the robot should pick it up, and into which pile to throw it in.

There are of course some things that remain unsorted, which we call residue: Something the robot is not supposed to pick. It should only pick wood, mineral, metal and plastic of a certain size: Everything else falls off from the end of the conveyor line into the reject bunker.

Are the robots fast enough to pick everything on the belt?
It depends on how fast we want to run the conveyor and the robots, because there are different modes. When there’s a lot of incoming waste, we can select a fast mode where the robot just picks the most valuable objects. If you run the conveyor slower, the robots can pick everything, but maybe the robots will spend idle time waiting for objects. So it’s all about efficiency: to select the optimal speed depending on how much waste the plant has at that particular time.

Is this optimization automated? Or do the technicians manually choose settings based on the volume of trash on a day-to-day basis?
The system has an automatic speed control that optimizes the speed. But you need to tell it how many tons per hour you want to process for it to maintain that flow. Since the robot cannot know how much waste gets in, someone has to manually input that information through a user interface.

Watching them, it seems they are very accurate most of the time.
Yes, they are. Recognition is one of our key technologies, the other is the picking. Moving these robots in real time and making them pick the right things is our specialty.

Picking is a very hard problem: As you can see, the shape of the object is completely random. Typically, robots repeat the same movements over and over again to learn. You can constrain the problem to make it easier, but since the objects here can be of any shape, we use artificial intelligence to determine from what position and angle is best to pick up an odd shaped object. And it works.

When they pause for a moment, are they “thinking”?
It depends on the situation. Sometimes when they pause there isn’t anything to pick. Also today we are testing a different software for the second robot, which might also cause more of these “thinking” pauses and affect the speed performance.

What is different in the new robot you are testing, and what are you trying to achieve?
We are all the time trying to develop new features for the product, it’s an ongoing process. Introducing new materials that make the robots faster, for example.

Who has been working on the intelligence for the sorting algorithm?
We have about twenty people on our R&D team working on this. It’s been a multi-year project. The company was founded in 2007 and in 2009, the development of this product really started. We produced the first commercial product in 2013 and then, during the last year, we made a big upgrade that improved speed by 50%. We also made the robots more durable, among other things. What you see today is the second version of the product.

How many of these robotic sorting systems have been sold so far?
For our version one product, we have shipped seven robots. For the second generation five have been sold so far, and we are actively selling more.

How do you reach target buyers? Finland is quite small, so how do you reach a global market?
Our focus is Europe at the moment, and actually all our sales people are traveling in Europe right now. We also take customers to visit our waste sorting facility in Helsinki to show how well it works. It’s good that customers see this is not just an idea in a lab, but an actual production unit that actually makes money for the customer. We have a lot of interest from abroad: visitors come here all the time.

What is your company’s plan for future developments?
Our current focus is now on these products. Products that really work, and are efficient. A lot of effort goes into selling these, of course. In R&D, there are certain aspects of these products that we want to improve, like adding more divisions of objects that the robots can pick. Currently they pick wood, mineral, stone, metal, and plastic, but there are other kinds of materials that would be valuable too, so we are adding those. We are constantly looking for new application areas. I can’t disclose anything concrete at the moment but, for example, areas in which unstructured objects need to be manipulated, such as mining, would be a good use case.

Talking about AI more generally now, the current hot topics on AI are about the future relationship between robots and humans. For example, in Asia, robots are already beginning to fill human roles, such as robot waitresses, nurses, and guards in jails. Do you see ZEN Robotics going beyond industrial applications and into social robotics?
Some of our technologies could be employed in that area, yes. But our focus is more on the industrial side, rather than on interaction with humans.

The discussion in media quite often turns to fear about the dangers of AI. Recently, prominent technologists like Bill Gates and Elon Musk have made very public statements warning about the possible dangers of developing AI without boundaries. How do you feel about this?
The Artificial Intelligence we have right now is very far from being dangerous. It’ an interesting academic discussion, but it’s not relevant yet… Maybe in the distant future. It’s a bit hard for the general public to understand how AI works: it’s not the kind of self-programming AI that people are imagining, that they fear. But I guess it’s better to be a bit conservative than to do something stupid.

Do you see a future where AI could become self-aware, and improve itself?
An AI that re-programs itself is still far away. But the point of AI is that it learns, with some support from the programmers. An example of this is how our AI is trained to pick objects: when children are learning how to pick up a mug, you don’t tell them: “grab it this way… no don’t touch that, do this”. Instead, you let them try for themselves: the child picks it up and… oops, it falls. They try again. Sometimes they succeed, sometimes they fail. And the robots learn like that: they try different ways, and see which ones work and which ones don’t. Then they try to repeat the picking methods that succeeded in the past, and avoid picking methods that failed. With trial and error, they learn. So yes, they improve themselves, but not in the way people think. For example, our robots can improve at sorting trash, but they can’t become self-aware. They are used for a very specific task and they learn to be better at that task only. Even if we wanted them to sort anything other than waste, that would require human intervention. It would be impossible for our trash-sorting robots to take over the world, for example!

Could a robot learn like a baby, from a blank slate, and develop its own rules and behavior? Should we discuss the ethics of something like this?
I think it might be possible for robots to learn specific things in that way… But to freely learn anything? Not in the near future that I can see. Of course, no one can tell what will happen in a hundred years, but I don’t think AIs will represent a threat to humans. Still, it’s valid to talk about it, to understand it and define it. The kind of intelligence we’re talking about, a distant intelligence from the future, will be much different from the intelligence in use today.

Popular fiction is filled with many apocalyptic and dystopian scenarios of robots and androids clashing with humans. In a real case example in Japan, a team of researchers was designing algorithms that replicate human emotions and behaviors for their androids. But in one situation, the android “liked” one of the researchers so much, that it blocked her from going out the door of the lab! She had to press a panic button to get help from the assistants. How real do you think are the dangers of flawed algorithms and wayward robots, as a threat to society?
Researchers in different labs use very different kinds of AI algorithms. But I think in general, it’s safe to say that the development of AI is still in the very early stages. We are still away from such scenarios happening outside research environments.

It fascinates people to think about robots that could behave like humans, don’t you think?
Yes, and also robots that can look “human-like”. If you look at our current robots, they don’t look like humans. Actually, we used to have a 6-axis robot that looked like a human arm and could move freely. So that was kind of “cooler”, but our current models are more efficient. This one is also better from a safety aspect, since it has four legs and can’t come outside that area no matter what it does. The older model was a six-axis robot standing on one leg. That’s why there are cages protecting the robots. You don’t actually need to be scared that the robot does something unexpected… Just about humans entering the area where the robot is. But with the older robot model, it was “what if it turns around, or if it does this or that?”.

Thanks so much for your time, and for letting us peek into the fascinating world of real-world robotics!
You’re welcome!

This is ZEN Robotics’s website.