Each day, the world produces more than 3.5 million tons of plastic and other solid waste, according to the World Bank. The U.S. produces 4.4 lb. of trash per person, per day. As public concern rises about non-biodegradable pollution in the oceans and food chain, interest has grown in recycling, but there are significant political and processing problems. For the latter challenges, researchers and startups such as AMP Robotics Corp. have offered solutions for automated waste collection and handling.
Trash has become a political issue, as landfills fill up and cities ship trash ever farther, often to the developing world. Last summer, China restricted imports of plastic waste, forcing the U.S. to look for alternatives. The Philippines and Canada had a diplomatic spat over tons of mixed trash that Canada had shipped to Asia years ago. Some municipalities and states have begun banning certain single-use items, such as drinking straws and grocery bags.
Without destinations capable of processing such materials in sufficient volumes, many local recycling programs ceased, sending tons of paper and plastic back to landfills or incinerators. Sorting recyclable materials from the trash stream fits the definition of “dull, dirty, or dangerous” work for which robots can be developed. AMP Robotics is among the startups that are applying automated efficiency to sorting a variety of recyclable materials.
The Louisville, Colo.-based company has been recognized by the National Waste & Recycling Association and the World Economic Forum as an innovator in sorting and data acquisition. Last month, AMP Robotics launched its AMP Cortex dual-robot system (DRS) for recovery of municipal solid waste, electronic waste, and construction and demolition waste.
Matanya Horowitz, CEO of AMP Robotics, recently spoke with The Robot Report about recycling robots guided by artificial intelligence.
What was the development timeline for the AMP Cortex?
Horowitz: From the beginning, we were sure to focus on the commercial aspects. We initially focused on construction and demolition waste, which have fewer materials. We focused on difficult-to-sort materials such as big objects or hazards. But we also talked with recycling facilities.
In 2015, we had our first demonstrations, and trade groups such as the Carton Council introduced us to others who were ready to work with us. Alpine Waste & Recycling was a great partner. It was willing to test our products in a closed loop and was patient in figuring out issues. By the end of 2015, we had switched over to consumer stuff like seasonal packaging because so many groups were ready to work with us.
For all of 2016 and 2017, we focused on picking core pilot customers. We started in Omaha and Minneapolis, where existing systems typically worked well but there were weird problems. Plastic bags got caught, and big pieces of wood trashed robots.
In 2018, we worked on reliability and started taking on more installations. AMP has over a dozen sites with its technology.
The AMP Cortex combines off-the-shelf robots with machine vision and AI in the cloud. Source: AMP Robotics
How much did AMP Robotics rely on off-the-shelf components?
Horowitz: Our thesis was that the hardware largely exists. We use off-the-shelf robots from ABB and Omron for manipulation.
What’s missing is the vision system, so we developed our own built on deep learning to do identification of materials. Our initial hope was to be entirely software.
We also developed specialized custom suction grippers, because they need to be highly reliable, low maintenance. We played with mechanical grippers and found that suction worked best for our needs.
This is coupled with a proprietary vision system. We initially looked at hardware partners, but they couldn’t deal with environmental hazards or dust.
There are a lot of robotic gripper makers out there. What design requirements did you have for your grippers?
Horowitz: There are some unique things about our industry that worked to our advantage. You don’t need to shoot for perfection [in grasping items] — 95% is acceptable for now. Stuff is already damaged, so if shredding something helps, we can do that.
In trash sorting and recycling, we don’t have the energy constraints of other industries. We can get as much vacuum as we need to pick up objects.
My Ph.D. is in robotic grasping, geometry, and robotic motion, and I had exposure through JPL [NASA’s Jet Propulsion Laboratory] projects. Robotic manipulation is very hard. I got out of it the lesson to do the simplest thing you can, overpower it, and don’t try to be too clever.
A lot of people were pretty skeptical initially about our approach. Some academic robotics guys thought hands were the way.
The AMP Cortex is a dual robot for recycling that combines proprietary AI and existing hardware. Source: AMP Robotics
What was your approach to machine vision for sorting?
Horowitz: If a person can do it by sight, we should be able to do it with vision and deep learning. AMP Robotics amassed a massive data set of bottles, cans, and more. We used it to teach systems to recognize No. 1 and No. 2 plastics such as milk jugs, sizes, colors, logos, text, and reflectivity — a whole set of features.
We incorporated all sorts of information and figured out what’s relevant and what’s not. We trained AMP Neuron on every type of material that the [recycling] facilities care about, to distinguish between aluminum foil, cans, and cat food tins.
They can decide what they want the robot to pick or not. That’s a core need of the technology — the robots must identify everything and learn from one another, even if they don’t pick everything. They can learn from facilities across the country.
Were there any concerns about data security?
Horowitz: We have good proprietary agreements with facilities. We make sure customers understand that an inherent part of the technology is getting better and being able to learn from the experience in your facility.
AMP is working with ERI as it scales up its groundbreaking electronics recycling.
The AMP Cortex can pick 160 pieces per minute. How did you get it to be that fast?
Horowitz: For the dual-robot setup, that’s 80 pieces for each of two robots in one enclosure. In a typical facility, there are opportunities for one- to two-dozen robots. There are also mechanical means of separating containers and paper, so each has different lines that are 30 ft. long and have six people. A facility could have 24 to 36 people, and we can automate three-fourths of locations with Cortex.
People can also do 80 picks per minute, but they get tired, and the industry average is closer to 40. There are already high rates of turnover.
There are still things for people to do in presort, taking out things like tricycles and engines. We don’t have robotic grippers for that.
How can robotics help change the economics of recycling?
Horowitz: Automation is absolutely part of the solution for trash handling. In the U.S., we already recycle water bottles and newspapers, but the challenge is that transport and sorting costs rise to commodity prices. Companies don’t want to bear that risk.
If we reduce the cost of sorting, that changes the economics of what materials are worth recycling. In many facilities, labor is 50% of operating expenses, and that affects the profit profile of these materials.
If you go around the country, there are lots of facilities that are doing OK despite low commodities prices. It depends on the contracts and the quality or purity of the materials they’re able to produce. Robots can maintain focus on the quality of materials, particularly for certain types of plastics and metals.
How is demand for recycling automation changing? Could your technology be useful for other markets?
Horowitz: A lot of facilities see automation as a way to drive costs lower, even in a depressed commodities market. We’re growing internationally. We’re right now in Japan and Canada, and we’re exploring opportunities in Europe, Australia, and OECD countries that have strong recycling infrastructures.
We stopped doing construction waste for a while, but we’re now getting back into it in Japan with Ryohshin. We needed data sets for construction and demo waste, but the tech stack was able to move over. Deep learning helps with consistency. We were also able to move over for electronic waste.
We’re aware of similar applications in agriculture, logistics, and manufacturing. We’re a little selective in choosing applications based on where does our technology work. We’re really focused on recycling, and we’re waiting for a good partner who knows a market well and can take advantage of our tech for other challenges.
There are a number of other recycling robots in development. How do you see the competition?
Horowitz: Some other groups are trying to do similar stuff, but it’s really hard to build reliable equipment that can perform in the materials stream. There’s high demand, but it’s pretty tough. Partners can help us expand and version the technology for other niches, such as electronic waste.
While competitors are focused on other pieces of capital equipment and multimillion-dollar retrofits, we have a modular design that can quickly scale.
What about collaborative robots that can work alongside people?
Horowitz: Collaborative robots are good for the goal of low retrofit. Our first iteration had cobots, but if customers have the [conveyor] belt space and need throughput per linear foot of belt, they need to sort more at a lower cost per pick. Cobots are exciting, but they haven’t yet seen the pick rates that customers ask for.
Speaking of people working with robots, we are hiring. We recently hired Brent Hildebrand as managing director of enterprise sales.
Filed Under: AI • machine learning, The Robot Report, Robotics • robotic grippers • end effectors
Tell Us What You Think!