Deep Learning Meets Trash: Amp Robotics’ Revolution in Materials Recovery

In this episode, Abate flew to Denver, Colorado, to get a behind-the-scenes look at the future of recycling with Joe Castagneri, the head of AI at Amp Robotics. With Materials Recovery Facilities (MRFs) processing a staggering 25 tons of trash per hour, robotic sorting is the clear long-term solution.

Recycling is a for-profit industry. When the margins don’t make sense, the items will not be recycled. This is why Amp’s mission to use robotics and AI to bring down the cost of recycling and increase the number of items that can be sorted for recycling is so impactful.

Joe Castagneri
Joe Castagneri graduated with his Master of Science in Applied Mathematics, with an undergrad degree in Physics. While still in university, he first joined the team at Amp Robotics in 2016 where he worked on Machine Learning models to identify recyclables in video streams of Trash in Materials Recovery Facilities (MRFs). Today, he is the Head of AI at Amp Robotics where he is changing the economics of recycling through automation.

transcript

---

[00:00:00]
(Edited for clarity)
Abate: Welcome to Robohub. Today, we’re in Denver, Colorado, speaking with Joe Castagneri, head of AI at Amp Robotics. It’s staggering how much trash materials recovery facilities (MRFs) process: 25 tons per hour. And yet, much of this is done manually. Amp Robotics believes robots are the future of this industry. Joe, how did you get involved with Amp Robotics?

Joe Castagneri: At 19, while studying applied math at CU Boulder, I met Matan Horowitz, the company’s founder. Amp Robotics was in its early stages, experimenting with sorting using an Xbox Kinect sensor. After seeing a presentation on robotics and recycling, I joined as an intern in 2016 and transitioned into machine learning by 2019.

Abate: Fascinating. So, the company’s foundation was built on AI?

Joe Castagneri: Exactly. The goal was to merge robotics, AI, and green tech to address major societal problems. Matan saw recycling as the right challenge for our tech.

Abate: Given the advances in GPU technology, did you begin with cloud processing?

Joe Castagneri: Actually, we opted for edge computing due to poor internet in trash facilities and the need for real-time operations. But as we grew, we shifted some support functions to Google Cloud.

Abate: How did Amp Robotics evolve from its early days to its current state?

Joe Castagneri: By listening and learning from our failures. Each robot deployed taught us valuable lessons. Rapid iteration and understanding customer needs were essential. The challenge lies in the diverse and unpredictable nature of waste.

Abate: Absolutely. Recycling facilities deal with so much variety in trash items.

Joe Castagneri: Indeed. Consider a milk jug; its appearance can vary greatly. Traditional computer vision struggles in this space. But deep learning, with enough data, can tackle this complexity.

Abate: And packaging materials and designs constantly evolve. How does the AI handle these changes?

Joe Castagneri: The key is consistent retraining and adaptation. Our models need to evolve as the industry and materials change. Model maintenance is crucial in this ever-shifting environment.

Abate: It sounds like this industry experiences significant model drift.

Joe Castagneri: Yes. Good way of concisely putting it. Totally agree.

Abate: So, and then here behind you, we have this, not a prototype, but like an in-assembly, model.

Joe Castagneri: Yes. So this is our flagship cortex product where we have a Delta style robot that will overhang over a belt. The belt will go from where I am through here. This unit in particular, we’re on our production floor where we manufacture the units we assemble. The robots that are Omron robots, we integrate with Omron and then we custom design the pneumatics and the wiring, the frame, the vision cabinet that is running that edge compute. And we bring it all together into one package. So this one is in process of manufacturing, and will go out into a recycling facility over a conveyor belt.

Abate: Yeah. So this is a five or six year old prototype called Claudia. So to explain, you have a suction cup gripper here and a beefy spring so that the variable height of the material or condition of the material is absorbed mechanically.

Joe Castagneri: And then a pneumatic system going through this particular gripper and the suction cup will form a vacuum seal and we descend, suck, and then place off the side of the belt into a chute or into a bunker.

Abate: So then this right here would be where, say a milk jug would come and it would hold onto that milk jug.

Joe Castagneri: Yes. It’s air suction and in particular, ahead of the robot cell, a camera imaging the conveyor belt will look at the material, localize where it is and what it is. And then the robotic path planning software will say, okay, I am configured to pick these things, so let me subset down what I’ve seen to what I’m configured to pick. Right. And then, there are too many things to pick that I have time for. I want to optimize the number of things that I can pick, given how long they’re gonna be in my picking region. And then I will intercept to be at this location at this time and turn my vacuum on at this time. And then place it off the side of the belt.

Abate: Yeah, so the interesting thing here is that this is a moving belt. You’ve got limited belt amount of time, and you’re trying to hit a certain number of items per minute that you’re picking.

Joe Castagneri: Yes. Right. In particular, the value proposition of these units is as a replacement for human sorters. And so human sorters will remove material at 30 to 50 picks per minute, at their peak. So a decent starting robot will remove material at 30 to 50 picks per minute to break even with a person, but really, you would like it to do better. And so these systems routinely hit 80 plus picks per minute. We’ve seen them hit over a hundred if the material stream is perfectly providing you a lot of eligible options in a well spread out way. So, a lot faster than a person, at a higher purity and for the whole duration of two shifts a day.

Abate: And how does that change from, say, one facility to another? Are these used in different ways by different companies?

Joe Castagneri: Dramatically. Yes. There’s always a conveyor belt in a facility. That’s the last chance Conveyor. And it’s the very last one. It’s your last chance to get any stuff on that conveyor or it’s gonna go to landfill. And this is a frustrating thing to consumers because you figure, you put it in your recycling bin, it’s all gonna be recycled. And the reality is, it’ll be passed through this facility and whatever the yield of that facility is, we’re gonna pull that out. The rest goes to landfill. And so our early applications were to put these units on last chance lines and hey, get whatever you can. But a different type of application for these might be you have other conventional sorting equipment that is separating 2D paper and cardboard from 3D containers and plastics, and you have all this paper and cardboard, but because it was sorted conventionally, there are a whole bunch of other things in there. And so you would quality control, remove stuff out of that stream. Historically, this has been done by people. If it’s not done, then the paper bales that you make might be rejected by the buyer. There’s too much plastic in there, too many impurities. So it has to be done to ensure that the product you’re making, paper in this case, has any value. And these can be there to quality control that stream.

Abate: Is it a mixture of everything that people put into their recycling bin is now what arrives at the MRF. And now you have to separate each individual component. So it would be like you’re separating out the paper, the plastic, the cans, and then the random trash that people threw in there as well.

Joe Castagneri: That’s exactly right. I go one step further. If you think about the waste stream, like a miner thinks about ore, what do you have in there? You’ve got precious metals, hydrocarbons, paper products, wood products, but the problem is they’re not refined. If you can sort them, you add value. It’s trash until we can sort it, and then it becomes valuable. This is a feedstock now. It’s no longer trash. It’s transformed into an input to an industry. So when people throw stuff in the recycling bin, they will wish cycle things, thinking, “Oh, I bet they’ll find a use for this.”

And it arrives at a recycling facility, dumped in a massive pile of recycling, and a front loader takes a scoop of it and puts it into the system. The first conveyor belt in the system is called the Presort line. It’s usually a really wide, rugged conveyor belt with hand sorters pulling off items like bicycles. This job is still done by people because it’s a difficult grasping problem. They remove really odd items that shouldn’t be there, like bowling balls, dog waste bags, bicycles, mattresses — things that can break machinery down the line.

Then, conventional sorting equipment sorts through it.

Abate: How does a mattress get into a recycling can?

Joe Castagneri: The recycling dumpsters in cities, typically. In my building, for example, we have a dumpster for garbage and one for single stream recycling. People will put their old Ikea lamp in there because it has metal. They think it’ll be recycled. But since waste is so abstracted away from everyday consumers, they don’t realize that these facilities have to run at 25 tons an hour to be profitable. They don’t have time to disassemble that lamp. It stands in the way of efficiency.

Abate: 25 tons an hour.
Joe Castagneri: That’s common for municipal facilities. In Denver, for instance, they might process 25 tons an hour, or 50,000 pounds an hour of material.

Abate: And do you know offhand how much trash a person produces in a year?

Joe Castagneri: I think a family household produces about three tons. About one ton of that is recyclable.

Abate: So this is on a massive scale.

Joe Castagneri: Absolutely. Trash is produced locally, so you need these facilities locally. They’re called municipal recycling facilities because they’re often funded through municipalities to support the local population. No city is the same. Denver, a big city, having a 25 ton per hour facility for recycling makes sense. In Colorado, if you go into the Rocky Mountains, it’s rare to recycle because there isn’t enough volume to make it profitable.

We’re concerned about why there isn’t recycling in more rural areas, or in areas that don’t have the population to drive 10 to 30 tons an hour of waste. You need enough volume for the business to be profitable. It’s a narrow margin, so you need scale. It would be great if we could build a smaller facility that was profitable without requiring so much throughput. That’s another thing we’re looking into.

Abate: So, what are those fixed costs that are preventing people?

Joe Castagneri: The fixed costs for a facility include the capital equipment, the sortation equipment, and conveyor belts. If you visit these facilities, it’s a maze of conveyor belts transferring throughout. Just considering the conveyor belts, they are a major expense. For instance, a facility processing 25 tons per hour might cost 10 to 20 million to build. In the mining industry, this might not seem like much, but in other sectors, it’s substantial. Given the thin margins on recycling, justifying that $20 million can be challenging. So, the primary fixed costs are the sortation equipment and the conveyor belts. Then there are dynamic costs, like sourcing material and paying for freight both to bring materials in and ship sorted goods out.

Abate: With tight margins in this industry, how much are operations affected by changes in material prices or varying regional prices for certain materials?

Joe Castagneri: It’s hugely impactful. For instance, in 2018, China stopped accepting low-grade plastics from the US. This was disruptive because instead of earning from these plastics, facilities had to pay to landfill them. This sparked a need for innovation, to find new uses and methods to handle these materials.

Abate: What counts as low-grade plastic? Bottles or items like plastic bags?

Joe Castagneri: Great question. The main valuable commodities in recycling are aluminum cans, cardboard, PET drinking water bottles, and HDPE milk jugs. However, there are other materials like colored HDPE and polypropylene, which also have value. Materials like polystyrene, used in red solo cups, are challenging to sort and don’t have as much value. When China stopped importing these low-grade plastics, the industry felt pressured to find new sorting methods and uses for them. It’s now leading to innovative techniques like pyrolysis and metalysis that can process these plastics.

Abate: With these valuable materials you’ve mentioned, are they primarily what your algorithms are trained on?

Joe Castagneri: Of course, there’s an incentive to be good at detecting and sorting the most valuable materials. However, AI robotics in recycling is also efficient at identifying materials that are typically ignored. We are part of the solution for materials that don’t have an established sorting process using conventional methods.

Now we are really adept at identifying the mainstay items of recycling because the robots came into existence when our company began retrofitting value into existing facilities. When retrofitting value, you need to accommodate the facilities as they are. They sort natural high-density polyethylene, PET bottles, cardboard, and aluminum, among others.

Abate: Okay. Because the MRF is selecting what they can sell, they’re choosing what their local customers are willing to buy. Some materials might not be valuable enough for them to pick. So, could they use the software to specify which items they’re interested in?

Joe Castagneri: Absolutely. They can configure what the robot will pick with just a few clicks. If halfway through the day they decide they want to pick a particular item from the conveyor because there’s more of it in the load, a few adjustments and it’s set to be picked. On the flip side, if they feel the machine is letting too many valuable items like PET bottles pass, they can increase its priority. These robots are highly adaptable, making them stand out in an environment where traditional sortation equipment is easy to operate but not versatile.

Using AI as the primary recognition tool in our facilities, we can change the type of material we’re processing and swiftly reconfigure the entire plant to adjust to the new material.

Abate: That’s quite powerful. Considering a system operated by humans, there’s a limit to how many items you can instruct them to recognize. Plus, switching tasks frequently can be disruptive. Has automation introduced notable benefits for your customers?

Joe Castagneri: Indeed. Hand sorting, for instance, epitomizes dull, dirty, and dangerous jobs. It’s risky due to hazards like needles and harmful substances in the trash. Workers wear protective gear, and the environment isn’t conducive for long hours. Automating this process proves advantageous. Our robots not only replace labor costs but also generate revenue. This leads to a return on investment in under two years for units like these. While humans might struggle with sorting a wide variety of items efficiently, AI doesn’t have this limitation.

Furthermore, there are other costs that aren’t immediately obvious. It’s challenging for a worker to keep multiple items in mind for sorting. Some data suggests that the average duration of employment for hand sorters is three to six weeks. The turnover can result in lost revenue, recruitment, training, and other associated costs. Automation proves invaluable in these contexts.

Joe Castagneri: Our biggest market is the United States primary sortation. We’ve installed more than 300 units in our facilities and in retrofit facilities that are operated by customers as well. Most of those are in the United States. We do have a small presence in Canada, Japan, and the EU as well. So we are international. Same problems exist in different markets. The EU has more regulatory pressure for solutions, leading to stricter purity constraints around the goods that you’re sorting.

Abate: And what’s that range? Is it like 95%?

Joe Castagneri: When we make bales of materials, big cubes of plastic, and sell them to a plastics reclaimer, the quality of that bale depends on if they hit the yield they were hoping for. If they didn’t hit the yield, then the bale was considered bad. Until now, we haven’t really known the exact contents of the bale. We assume it’s about this pure, but that’s a rough estimate. A rule of thumb has been for plastic bales, you want them to be 85% pure. For aluminum cans, you want them to be more like 97% pure. The reality is that recycling has historically been about doing the best you can, providing feedstocks to downstream processes and hoping they can work with the quality of material they receive. The EU is tightening regulations by requiring more recycling, even of low-quality plastics not often recycled in America.

Abate: So it’s not just about recycling more cans and bottles but also recycling more types of materials?

Joe Castagneri: Exactly, yes. You want to optimize both aspects.

Abate: But how can you start recycling more materials until you have the buyer side of the equation sorted? Like, is that sorted for them already? Do they already have customers lined up to buy these materials?

Joe Castagneri: Part of it is, and since there are several links in the chain, who’s the buyer for you?

Abate: From what I understand, the buyer is the entity purchasing the packed material from the MRF.

Joe Castagneri: Absolutely. The buyer side would benefit greatly from a transparent market where different commodities are priced based on their quality. Right now, the market operates on a contract-by-contract basis. Buyers in specific regions tend to buy from known partners who have historically provided good quality material. If we had a more structured marketplace, more entrants could participate, identifying valuable commodities and accessing them without needing a web of personal relationships.

Abate: Do you even have a reliable way of determining the yield of each bale?

Joe Castagneri: It depends on the process. For processes like aluminum can recycling, you can weigh the bale before and after processing to get a mass yield. We typically have decent yield numbers, but they cover the entire operation. With the addition of AI analytics, you gain deeper insights, such as the efficiency of a particular unit or piece of equipment.

Abate: That’s intriguing. It seems like a significant differentiator for places without this system. One of the biggest challenges in waste management appears to be the lack of access to quality data.

Joe Castagneri: Yes. The data is invaluable to us. We can adjust the AI to keep up with changes in the waste stream. Moreover, in our facilities equipped with multiple vision systems, the key idea is using perception to drive efficiency. This approach results in better yields and the ability to recycle a wider variety of materials.

Abate: If you were to envision a smaller version of this system for a minor municipality, what would it resemble?

Joe Castagneri: Imagine a shipping container with a conveyor belt. Items are sorted using a pneumatic-based optical sorter. It’s a simple setup that could be used temporarily, like at music festivals. For rural communities, you might need something between that and a full-scale recycling facility.

Abate: So, in essence, it’s an operation without human intervention, other than someone loading the waste?

Joe Castagneri: Yes. Someone loads, removes, and configures.

Abate: Fantastic. Let’s go take a look.

Joe Castagneri: Certainly.

---

transcript

tags: Actuation, c-Industrial-Automation, cx-Industrial-Automation, Industrial Automation, interview, podcast, Robotics technology, startup

---