They are coming—for your trash. Sorting through 67 million tons of glass, plastic and paper is dirty, low-paid, mind-numbing work. Matanya Horowitz’s AMP Robotics wants to take humans off the job.
At an RDS of Virginia recycling center in Roanoke, two spider-like, 300-pound robots sort through an unending line of trash. One robot’s skinny leg, which relies on computer vision to detect recyclables, plucks a hunk of blue plastic off a conveyor belt, while the other’s grabs a piece of an old water bottle. The machine then places those bits into sorting bins using a vacuum gripper.
For the nation’s 600-plus recycling facilities, which process some 67 million tons of waste, these leggy robots from AMP Robotics are one answer to the current bottlenecks facing the industry. Even before Covid-19 struck, AMP Robotics was starting to gain traction. But as boxes from home deliveries piled up at recycling centers and hiring—already a tough proposition—got even tougher as workers feared getting ill, AMP’s business boomed. “It’s repetitive and not ergonomic, and you are surrounded by unsanitary stuff like hypodermic needles,” says AMP founder and CEO Matanya Horowitz. “With Covid on top of that, people are saying, ‘Do I really want to put my hands in this material that maybe came from an infected person’s house?’”
AMP, which is based in Louisville, Colorado, has sold or leased 100 of its AI-powered robots since 2017 to more than 40 recycling facilities in North America, Europe and Japan. They’re not cheap, at a cost of up to $300,000 (or around $6,000 a month to lease), but those recycling centers are betting that the hefty capital expense will pay off with lower employment costs and higher efficiency. Forbes estimates that AMP’s revenue this year will reach $20 million, double its $10 million for 2019. And there’s lots of room for growth: Recycling is a $6.2 billion (revenue) market in the U.S., and while the overall market has been growing at less than 2% a year (and declined this year due to Covid-19), facilities are trying to figure out how to get more out of their waste, the majority of which still ends up in landfills.
Thanks to both its technological promise and fast growth, AMP was featured on both Forbes’ AI 50 list of artificial intelligence companies to watch and the Survivors and Thrivers list of 25 small business standouts that outperformed during the pandemic. While Horowitz, who has a Ph.D. in robotics from Caltech, bootstrapped the business to start, AMP has now raised $23 million in venture funding. Forbes estimates it reached a $100 million valuation with its latest round, in November 2019, led by top VC firm Sequoia.
“In my opinion, Matanya is one of the 50 best roboticists in the world,” says Sequoia partner Shaun Maguire, a classmate of Horowitz’s who led the funding round.
A bearded and balding man who is just 33, Horowitz speaks with childlike wonder when talking about automation. He fell in love with robots watching Transformers and Voltron cartoons growing up in Boulder, Colorado. His father, Isaac, an engineering professor at the University of Colorado, was renowned in the field for his work on control theory, a concept based on the math behind planning and reacting to one’s environment that was used at the time for fighter aircrafts and chemical plants. The elder Horowitz (who died in 2005) did not talk about his work at home, but years later, his son received a Ph.D. in the same subject because he thought the underlying math would be necessary for robots and artificial intelligence. “It might’ve been genetic,” Matanya says.
Horowitz fast-tracked his education with college-level classes during middle school, and completed both bachelor’s and master’s degrees in electrical engineering from the University of Colorado Boulder (along with three other diplomas in computer science, applied mathematics and economics) in four years. He spent time in a research lab hacking Roombas to perform cooperative activities, like moving a chair together. After college, he moved to California, completing a Ph.D. focused on robotic path planning at Caltech.
While there, Horowitz was captivated by the emerging technology of deep learning, which can allow robots to see things in the same ways as humans through computer vision. While other young roboticists pursued flashier technology such as self-driving cars or autonomous drones, Horowitz noticed the drab field of recycling. He realized that the industry, slow to adopt new technologies, presented a ripe opportunity to use technology on the cutting edge without having to compete with Google or Lockheed Martin. He was also drawn to using robotics to decrease the amount of waste that isn’t recycled, and improve the environment.
To learn about recycling, Horowitz spent his weekends driving to recycling facilities around Los Angeles. During these jaunts, he was struck by how much labor was required to sort through the junk—and how unpleasant it was. “There’s a pile of garbage with diapers sitting on top, and somebody’s trying to drag a garden hose out of this thing and stuff is landing on them,” he recalls. That pain, he figured, made it an industry waiting to be brought into the modern era with automation.
In the fall of 2014, Horowitz quit his postdoctoral research at Caltech and moved back to his native Colorado to start AMP Robotics, named for its technological aim of “autonomous manipulation and perception.” By then, recycling robots had “occurred to pretty much everybody,” Horowitz says. But the idea could only become reality after faster computer processing speeds opened up the potential of deep learning to transform existing automotive manufacturing robots for a range of new conveyor belt-based use cases.
Recycling robots were further along in Europe than in the United States. Appingedam, Netherlands-based Bollegraaf, which had filed patents for recycling robots in the 1990s, for example, had created robots that rely on spectroscopy and the heights of objects to sort waste. Meanwhile, Helsinki, Finland-based ZenRobotics had ones that could sort construction and demolition materials with color-coding cameras, laser sensors and metal detectors based on neurorobotics research. Horowitz took a similar approach, but focused on single-stream recycling, where newspaper, cardboard and plastics are all mixed together, as the area to apply deep learning that could teach the robots how to recognize objects based on colors, shapes, textures and logos.
For the first two years, Horowitz and his team of fewer than 10 employees tried to build a working robot using limited grant funding, much of it from the government. They built early robots with wood and garden hoses instead of industry-standard materials like aluminum, which resulted in cheap machines that weighed just 35 pounds. In one case, one of his lightweight robots ripped apart after too many paper dust particles accumulated inside its machinery. Horowitz even planned to repurpose a treadmill as a conveyor belt, and was improbably saved thanks to funds courtesy of a nearby environmentally-minded brewery, Oskar Blues, that was giving out grants for recycling drives and for schools to buy bins. “We should have just gone for the beefiest robot we could from the beginning,” Horowitz says.
By the end of 2016, the team had found a durable solution with the spider-shaped robotics system, which weighed 1,800 pounds including the frame. But they continued to tinker with the sorting technology—and to find bugs. One week prior to shipping the first unit to a customer in Nebraska, the robot was still struggling to pick up plastic milk jugs. Normally, the robotic arm would punch the bottle to flatten it, and its cap would fly off. Once every 500,000 milk jugs, a frequency that translates to once every two weeks, the mechanism would fail. “The milk jug would fight back,” Horowitz recalls. “What would happen is the robot would smash into it, and the robot’s arms would just blow off.”
Horowitz took charge of coming up with a better way to apply the machine’s suction grip to the plastic material. The trick was to modulate the robot’s arm to speed up and slow down in a specific way, and to adjust the elasticity of its springs so that it would hit the milk jugs like a pillow rather than a piece of cement. Within a week, Horowitz had a new prototype. “He closed that gap in a week because he is probably 10 times smarter than everybody,” says Rob Writz, AMP’s business development director and employee number number six.
Horowitz and his team finally got all the bugs worked out of the computer vision in 2017. He took the robot to venture capitalists and raised $3.2 million that October, allowing the company to finally purchase high-quality equipment. He sold his first five robots that year. The next year, AMP began providing its artificial intelligence to Canadian firm Machinex, a leader in recycling equipment, for its robots in order to get its name out in the market quickly.
Sales began to pick up last year. Sarasota, Florida-based Single Stream Recyclers bought six robots in May 2019, then ordered eight more in September that year. AMP’s robots helped the facility cut operating costs and allowed it to take on a second load of recyclables to sort through. Also in 2019, RDS of Virginia, based in Portsmouth, signed a five-year lease on four robots, two to separate materials and two for quality control. Even before the pandemic, “we could not find anybody to come and work for us,” says RDS president Joe Benedetto. “We were paying well above minimum wage, and yet we just couldn’t find people—and when we did, they would work for a week and quit.”
While AMP’s technology is nascent, the economics of its robots make sense for most recycling facilities in the United States, according to recycling consultant Juri Freeman. Its machines cost up to $300,000 upfront, but recycling facility operators say they expect them to last five to ten years, and perhaps even longer. Although labor costs vary by region, recycling workers make around $25,000 per year. But the robots are far more productive than humans, with an ability to pick up 80 pieces of material per minute versus 40, so each machine can handle the work of at least two employees, while freeing those workers to do other jobs at the recycling center. Add in other employment costs, such as training expenses, workers’ comp and PPE, and just two employees could cost a facility at least $70,000 a year, meaning the hefty price tag for the robots should be repaid within three to four years.
That’s a long-term bet, but there’s an added plus: Because the robots are more accurate at sorting waste than people are, the recycling facilities, which operate on low profit margins, may be able to improve the amount of recycled material they sell. That’s especially important today as recycling facilities are still struggling to find new markets for their materials after China, the world’s biggest market, essentially banned almost all imported waste in 2018, a decision that resulted in recycled material that previously could have been sold ending up in landfills for not being pure enough. In fact, the rate of recycled materials in the U.S. stands at a mere 35%, according to the Environmental Protection Agency. “That’s where AMP comes in,” says consultant Freeman. “It can help the [facilities] get the cleanest, best streams. AMP is making a lot of headway cleaning up the end bales that are sold and getting them to be sold for a higher value.”
This year, AMP’s sales have picked up rapidly due to Covid-19. Not only have the sorting jobs become even less appealing, but the surge in online shopping has placed billions of cardboard boxes from Amazon and elsewhere into recycling bins. “Recycling facilities are looking at placing larger orders now,” Horowitz says. “Instead of one robot, they’re saying, ‘Maybe we need six out of the gate.’”
With increased demand and cash to spend, AMP is now focused on expanding outside of North America and increasing use cases for its computer vision system. It has begun to apply its municipal recycling solution to adjacent industries, such as separating batteries from wires in electronic waste or sorting demolition debris for construction. The goal: to build up a base of knowledge so that the robots are one day able to process just about any type of waste.
Horowitz envisions his robots turning landfill trash into useful products, much the way that Plains Indians utilized the meat, skins and even dung of each buffalo they killed. “You have all this material that society produces—plastic bottles, pieces of wood, drywall—and people pay for it, but then somehow it has no value once it’s in the dumpster,” he says. “Why aren’t we using every part of buffalo?”