AMP Robotics (Autonomous Manipulation and Perception) is an industrial artificial intelligence (AI) and robotics company automates the identification, sorting and processing of complex waste streams. AMP’s solution is a combination of state of the art computer vision, machine learning and robotics, that can identify and rapidly pick recyclable materials off a conveyor belt for recovery. AMP’s Neuron Vision System constituent learns and recognizes material in the dusty, commingled conditions of recycling facilities, while Cortex Robotic Sorting component picks and moves material. AMP has received numerous awards and gained international recognition, including The Circulars 2018 Award for “Circular Economy Top Tech Disruptor” at the World Economic Forum in Davos, and the NWRA’s (National Waste and Recycling Association) “2017 Innovator of the Year” award.

Matanya Horowitz, CEO of AMP Robotics, founded the company in 2014 with the mission of changing the fundamental economics of recycling. In the Fall of 2019, AMP raised $16m in Series A funding, led by Sequoia Capital. Horowitz developed and commercialized AMP’s AI platform, AMP Neuron, along with AMP Cortex, which automates high-speed identification, sorting, picking and processing of material streams.

Horowitz was individually recognized as Waste360’s “2019 Innovator of the Year” in addition to being named in their “40 under 40” list. He also holds a PhD in control and dynamical systems from the California Institute of Technology, with publications and research in control theory, AI, robotic path planning and computer vision. He was recently interviewed by Joanne Pransky for Industrial Robot. That interview is reprinted below, with minor modifications. The original interview can be found HERE.

---

Joanne Pransky: What was the first robotic project you ever worked on?

Matanya Horowitz: I don’t know if this really counts, but when I was about five years old, I had the idea that I was going to make a Transformer and I spent a lot of time sketching this robot with one of my friends. We didn’t ever really make anything; we just talked and drew a lot.

For most of my life, though I was always interested in robotics, I didn’t actually work in it because it seemed too science fiction and not realistic as a career. However, that really changed when I saw the DARPA Grand Challenge with those robots racing across the desert autonomously. I thought the technology was extremely impressive and a lot nearer than most people appreciated.

After that, I changed my focus to robotics and my first real brush with robotics was as an undergraduate of Professor Todd Murphey research lab at the University of Colorado Boulder. He had me hack some Roombas and just have them obey different control policies. I immediately had some early exposure to just how tough it was for robots to work. The theory was fairly straightforward but the act of getting the robots to follow these control laws, communicate wirelessly to another, deal with their actuators, etc., was impossibly hard. I couldn’t believe even with all my programming skills how clunky the demo was. So my first real robot experience, when I was an undergrad, was a tough one.

After undergrad, I pursued a PhD in robotic path planning, nonlinear and stochastic control, computer vision, and motion planning. After I received my PhD, I looked at various industrial applications in which I could apply these technologies. One of them was the recycling industry.

I was not the first to think about robots and recycling. I found papers from the ‘60 s and ‘70 s saying it should be automated and I found patents from the ‘90 s where people were trying to do this. What I saw was that existing vision systems simply couldn’t meet the industry’s needs unless there was consistency, in which for example, template based matchers could be used if every aluminum can in the recycling facility looked exactly the same. The big hurdle to cross here was that in recycling, no two pieces of material are going to look the same. They’re always going to be smashed differently, lit differently, have a different orientation, etc., and those issues needed to be dealt with head-on. I felt I could solve these unique problems with deep learning and cost-effective robotics.

Pransky: What type of robots does AMP use and how many systems has AMP deployed?

Horowitz: AMP uses International Protection Marking (IP) 54 delta robots from ABB and Omron (Figure 1). Delta robots are some of the fastest, cheapest and lightest robots for pick and place. They’re very easy to repair. If there’s some sort of issue with a robot, and in these recycling facilities you never know what’s going to happen, you just pop the arms on. And other groups in the industry are following our lead with delta robots.

We actually first deployed a Bosch cartesian robot. We broke that and then we deployed an ABB SCARA that was not dust proof. The dust basically got into the screw shaft and jammed it up. It was too bad since we had it working for months. Dust proof is essential. We learned that the hard way.

At this point, over four dozen robots have been sold, mostly all in the United States, with a couple also being sold in Canada and Japan (Figure 2). We also have a fantastic partnership with Japan’s Ryohshin. We’re about to begin deploying quite a few robots with them, which we’re incredibly excited about. The sales and installation velocity are only accelerating, so there will be more and more of these (installations).

Pransky: Vacuum suction looks to be very versatile as a gripper. Do you have alternative grippers such as magnetic and conventional jaw grippers, and when are these the best choices?

Horowitz: Vacuum systems are very powerful and fairly generic. We’ve been able to pick up a wide variety of materials – bottles, cans, paper, cardboard, wood, brick, and more, but vacuum is not a solution for everything (Figure 3). We do have magnetic grippers to pick up ferrous material. We’ve played around with different kinds of physical grippers; things that look like spatulas, things that look like fingers, etc.

The challenge is that in the recycling industry, material is non-singulated, and it can be hard to actually create a grasp plan quickly that’s not going to interfere with some of the neighboring objects. Suction and magnets are really nice because you get to follow a path overhead where it’s really clear what you’re seeing, and then just grab a hold of something and pull it out. With suction, you can actually dig out material really easily that may be half covered by other things.

When you watch humans manipulate the material to pick them up, they do all sorts of fancy things like press their thumb down, and that’ll lift up the far edge of the material and then they can grab the piece of material. We really just don’t have the sophistication do that with robotics yet. Whenever we’ve tried a fingered gripper or something like that, we’ve just not been able to beat the performance with suction. But we continue to play with all sorts of new and different (types of) grippers including other suction-based grippers, and other technologies such as adhesives (Figure 4).

Pransky: AMP created their own vision system. Did they ever approach an existing vision company with their requirements?

Horowitz: My perception, when we were starting the company in 2014, was that this opportunity wouldn’t be there if there already was an available vision system that could solve this problem. The facilities absolutely wanted robots and I kept asking myself why there aren’t robots already. There should have been robots here 10 or 20 years ago. The fact that the facilities were not automated meant there must not have been anything suitable. With deep learning, we were able to create the solution the industry had been waiting for (Figure 5).

---

One application that we run into is party cups, like Solo cups. The clear ones might be made out of polyvinyl chloride (PVC), polypropylene, or polyethylene terephthalate (PET). These three plastics need to be separated and not processed together, but visibly they all look exactly the same. Infrared can tell the difference between them.

---

Pransky: What has been the most difficult object to teach your vision system and why?

Horowitz: Things that look pretty similar visually. Some of the more tricky ones are objects that have the same kind of branding on them. For instance, there’s Pepsi plastic bottles and Pepsi cans. They have the same logos and same colors and initially the system would get confused between those type of things that have the same branding on them. With enough data we were able to overcome that. But one other unusual item that we’ll sometimes get is a magazine ad which has Pepsi bottles in the ad, and then the system will identify that as a Pepsi bottle and go to grab it thinking it’s a bottle. For the most part, anything that people have trouble distinguishing, the system will have trouble distinguishing, and those require a lot of data to overcome.

Pransky: You look to rely on normal vision sensors covering the visible spectrum – any benefits of using UV or IR sensors?

Horowitz: There absolutely are. There are certain materials that you can’t tell the difference between in the visible spectrum. One application that we run into is party cups, like Solo cups. The clear ones might be made out of polyvinyl chloride (PVC), polypropylene, or polyethylene terephthalate (PET). These three plastics need to be separated and not processed together, but visibly they all look exactly the same. Infrared can tell the difference between them.

We’ve decided to focus on areas where people are being used right now to sort the material, with the idea that if people can see and tell the difference between the material, then our vision system can too, and we can go after these existing applications where there’s not enough manual labor already. Once we’ve sort of taken on most of those tasks, we’re likely to add additional sensors.

Pransky: Is there anything that manufacturers can do to make the sorting of waste products easier?

Horowitz: There’s lots they can do but it’s hard to build consensus around these things. They look at us and we’re just a very small player in this big industry, so we can’t really count on them to do what we want. We’ve taken the perspective that we have to have an adaptable system that we can quickly reconfigure or can quickly learn the material stream and we can adapt to what other people do.

Fortunately or unfortunately, I don’t think it’s likely the packaging manufacturers will adapt what they do to help our robots, but that’s okay. It gives us a challenge. The robots get better over time, and to a large extent we can adapt the robots to the material. It’s a double-edged sword. If it was easier, then it would make it easier for other people to get in the game, and the harder it is, the harder it is for anyone who’s trying to catch us.

Pransky: Can any pre-sorting be done based on mass and density, e.g. floatation separation, air blast separation?

Horowitz: Absolutely and those techniques are commonly used in the industry but are fairly low accuracy. What you find if you use an air separator or a tool like an air knife, is that a lot of the paper gets caught with the plastic bags. If you use a float tank, you’ll find different types of wood that you don’t want together and different plastics will come attached to it. The result is that you have heavy machinery – screens, trommels – and all that machinery is surrounded by people who effectively clean up the mistakes those systems make.

That’s where the robots are used, so we actually do work in conjunction with those pieces of machinery and at this point, all of our installations are working with machinery like that. In some cases, it’s essential. One of the pieces of equipment is called an OCC screen. The OCC stands for old corrugated cardboard, and these big spinning disks with large gaps between the cardboard bounce over the top of the disks and other things fall between the disks, and it effectively separates out the cardboard from things that aren’t so large and flat. That’s sort of an essential pre-sorting step that supports the robots because now the robots don’t have to pick up these big pieces of cardboard that might be as big as a TV or similar.

Pransky: Can you talk about your business plan for the next ten years?

Horowitz: There’s a couple of angles to it. One is we’re simply expanding the geographic scope of our robot sales to Europe, Southeast Asia, and Australia. It’s a lot of work too, though, because we need partners in these areas and those partners have to hold (to) the same kind of standards we do for service, etc. Additionally we’re branching out into other technologies that can help the recycling industry. One of these is deploying our vision system without the robots for data collection in recycling facilities.

The facilities right now can’t count the number of bottles and cans that come in and they have to be very reactive. They can’t detect when a hazard comes in. What we can do is say, “There’s a slug of pop bottles coming through, 5 per cent more recyclables that are going to landfill than should be.” We can also say, “That’s a propane canister, that’s a hypodermic needle”, and stop the system for somebody to take care of this before somebody gets hurt. Our vision system helps them with safety and operations.

Pransky: What is the greatest challenge your company is facing today?

Horowitz: Things are going well. Our install base is growing quickly, and I would say it’s nothing existential for us, but it’s hard work to build a team for the long-term that can have international scope. That’s the task we’re engaged in for the next year – building a sustainable business behind the technology we build.

It’s challenging enough to build these robots and have them be best in class and extremely reliable. The first stage was, ‘Will it work?” We got it working. Then, “Can we get a couple out there where people buy it? Oh my gosh, a lot of people are buying it.” Now we’ve stretched ourselves pretty thin and need to catch the company up to the commercial success we’ve had.

---

Right now society needs a lot of these robotic solutions. So the other thing I’d say is that people should pursue a business opportunity if they have a good idea.

---

Pransky: In four years, you received four bachelor degrees and one Masters, and then went on to get your PhD. Did any of these prepare you for running a company?

Horowitz: I would say my training up to now did not prepare me for it in the slightest. I would say I overdid it when it came to the academic knowledge that was necessary. I thought I had to really know these systems inside and out to really be able to create them. That’s not true as long as you have engineers and programmers that you can really trust.

The only time my knowledge comes in handy is when I don’t think we should work on something, or if I think another direction for the technology would be better. But I don’t do that very often. We have excellent programmers and developers and so for the most part they come with great ideas and I just say, “Sounds good to me.” In the past couple of years, I had to learn how to run a business and I’d say that has not had too much overlap with all those degrees.

Pransky: Can you share the greatest lesson that you’ve learned or the biggest mistake you’ve made?

Horowitz: There would be a couple. For the first two years, we were fortunate to win some government grants. We won a National Science Foundation Small Business Innovation Research (program grant). We won a similar grant through the state of Colorado, called the Advanced Industries Accelerator Grant. These supported us for the first two years and we were actually hoping to get the company off the ground with them. But we were just a handful of guys moving too slowly. We had to be really scrappy. We weren’t paying ourselves much. I wasn’t paying myself for a long time and basically lived off of my wife.

In 2017, I just thought we were moving too slowly, so I raised venture capital. I was pretty wary and didn’t know what to expect from venture capitalists. The money is very helpful and allows us to reduce the burden on everybody by hiring, and gives our customers the assurance that the company will be around.

We were very fortunate to have investors who have been fantastic partners and have helped guide me in building the company. I’m really blessed to have that kind of support. They are very aligned with what we want to do. Many investors do not want to invest in recycling waste as it’s an old school industry. There are big players, incumbents, and it doesn’t move fast. There are lots of reasons venture capitalists gave me why they didn’t want to invest. The ones who did stick around and who wanted to invest were those who believed that our technology could be influential and that together we could build a highly successful business that can help advance a circular economy focused on eliminating waste.

Pransky: What is your proudest moment?

Horowitz: The moment we got the system first working in the facilities. We got the system working in our offices here fairly early on. We deployed it in a Denver facility called Alpine and it was a disaster. It didn’t work and we were breaking the system all the time. There were these other human sorters who were right next to us and saw us struggling and working there 16 hours a day. It wasn’t clear if it was going to work and you have these moments of despair. When it finally worked, the sorters next to us yelled, “Good for you!” Everyone was excited.

I’d also say that there are a lot of small moments in the company where you feel very proud to see people who are picking up the ball and running with it doing something amazing we didn’t know was possible. It’s just wonderful to see the members of the team internalize the company’s mission.

Pransky: What do you think PhD and masters of engineering students should be doing while in school to best prepare them for the commercial side of robotics?

Horowitz: A startup requires several skills. If you’re studying for a PhD or Masters, the assumption is going to be by everybody that you mostly have to make the right decisions on the technology. The unfortunate truth is that you have to be right about a lot of your decisions. There’s no way around it; you can’t be wrong that often. You need to have enough knowledge to be right about what is possible and what’s not possible, and how hard it’s going to be to make something.

Beyond that, it’s very important that you have people you’re going to be able to attract into the company. Meeting smart people who you will be able to track later down the line ends up being really helpful. That was under appreciated by me and it made recruiting at the beginning of the company difficult. Depending on where you go to school, that can have a big impact.

Right now society needs a lot of these robotic solutions. So the other thing I’d say is that people should pursue a business opportunity if they have a good idea. It’s not that risky and investors will invest enough money that you can pay yourself. There’s a lot of funding available to do these kinds of things. It’s an amazing time to just jump in and learn the skills you need to whether it’s management, finance or whatever. I encourage anybody to do that, but you do have to have some meaningful contribution. You have to be creating a product that people love; that makes an impact. There are lots of resources that are available now whether it’s Y Combinator or other organizations that can inform you on how to think about these problems. Those tools are wonderful and they’ll let even somebody who’s an outsider to startups get started, the way I did.

---

Joanne Pransky has been an Associate Editor for Industrial Robot Journal since 1995. She was also one of the co- founders and the Director of Marketing of the world’s first medical robotics journal, The International Journal of Medical Robotics and Computer Assisted Surgery. Pransky also served as the Senior Sales and Marketing Executive for Sankyo Robotics, a world-leading manufacturer of industrial robot systems. She has consulted for some of the industry’s top robotic and entertainment organizations, including Robotic Industries Association, Motoman, Stäubli, KUKA Robotics, ST Robotics, DreamWorks, Warner Bros., and for Summit Entertainment’s film Ender’s Game, in which she brought never-seen-before medical robots to the big screen. She can be contacted at joannepransky[@]gmail.com.