AMHERST — Peanut butter jars, takeout containers and soft plastic wrap often end up in the recycling bin, contaminating viable plastic, cardboard and paper for recycling and resulting in more garbage in landfills.

But if consumers confused about recycling rules are instigating this contamination problem, Somervillie-based robotics startup rStream believes they can solve it by using artificial intelligence to sort trash and recycling instead.

“Our objective is to make equipment that can sort of waste and recyclables,” said Ian Goodine, co-founder of rStream Ian Goodine. “That could be to service the function of separating single stream into various commodities that can ultimately be packaged and sold for remanufacturing. The other function which we’re doing here today is actually pre-sorting.”

Last week, rStream piloted its automated trash-sorting AI at the UMass dining commons and the UMass Waste Recovery and Transfer Facility, practicing its identification and separation of plastic, paper and cardboard. The more data to fine-tune its  algorithm, the better rStream’s robot technology can produce a cleaner recycled material for waste companies to sell, leading to a better recyclable product and lowering the amount of waste sent to landfills each year.

“We’ve all experienced that feeling of staring at the waste bin, not knowing where to put the thing,” Goodine said. “Multiply that by 50,000 people every day, and there’s a lot of missed opportunities for the sustainable recovery of those materials.”

Using artificial intelligence’s ability to analyze photographs, rStream’s robot views each piece of trash inserted into the trailer through a camera, and attempts to match that image with other depictions of trash in its dataset. If the computer correlates an item with one of the recyclable materials in its database, it redirects it to be recycled.

“It’s building the capacity to work with individual customers to identify what their specific sorting needs are,” said Kathy Wicks, dining sustainability director. “Right now, it’s set up for recycling and trash, but eventually we could separate out all the UMass water bottles, and send them off to a company that recycles those bottles into clothing that we then sell at the UMass.”

This week has been a homecoming for Goodine and his co-founder Ethan Walko, who conceived rStream as undergrads at the UMass College of Engineering. Walko said he and Goodine worked in a polymer science lab investigating chemical processes for recycling polypropylene, a hard plastic used in chairs, crates and bottles. However, they had a difficult time sourcing polypropylene from the dining hall despite its frequent use.

“As we kept going down this rabbit hole and started talking 100 different people that work in the waste management industry, we started to see all these lines are pointing at this whole collection challenge that exists with trying to get all the recyclables in recycling bins separate from contamination,” Walko said. “It’s a monster of a challenge that, once we heard it a bunch of times, we couldn’t not try to do something about it.”

Since graduating UMass in 2022, Walko and Goodine have raised $3 million for their startup and became Activate Fellows to further develop their concept. The datasets used to create artificial intelligence are extensive, as not only did the entrepreneurs teach the computer what recyclable materials are, but all the various crumbled, crushed and crinkled forms it comes in.

“With hundreds of thousands of photos, you can build this model that’s good enough to sort of all the different types and variations of trash, because trash is really challenging,” Walko said. “Sometimes the water bottle label is gone, sometimes the cap is missing. So all these things, you have to become really robust, which is why it takes a lot of data and a lot of time and skill to make something that works.”

The biggest challenge rStream faced wasn’t building the AI model, but training it. Walko recalls the first time they tested their program at their home base in Greentown Labs in Somerville, wading through foul-smelling trash in waste removal suits. Eventually, Walko and Goodine went searching for larger quantities of garbage. Without a way to bring trash to their computer, they decided to bring the robot sorter to the trash.

“Having it be mobile, so that we can bring it to the site where the waste is generated, instead of trying to bring the waste to our lab, has allowed us to get the technology out of the lab much sooner than we previously thought,” Walko said.

Walko and Goodine have seen their AI improve in accuracy with each test, not only separating trash from recycling but sorting different recyclable materials to bale and sell. The AI has even learned about edge cases like the dreaded peanut butter jar or ketchup bottle, which often cannot be recycled due to food residue.

“Part of our core technology is to have a sufficiently easy to train model, because we’re dealing with waste, where it continues to change, and even within the known variety, you get all sorts of things you’re not prepared for,” Goodine said.

However, the key to the AI model’s success isn’t the technology or the trash, but the waste management workers who guide Walko and Goodine. Wicks said they’ve been very receptive of feedback, listening to the needs of the industry and expanding their model to be customizable for different sites.

“It allows us to dynamically decide, based on the rule set at this specific venue,” Goodine said. “Different haulers, different buildings have different recycling rules, os we get to find out exactly what’s there, and you work with people like Kathy and the UMass sustainability team to map those detections to the roots.”

This week, rStream has sorted through trash at various buildings, from the Campus Center to the Mullins Center, to learn more about different scales and situations where their trash trailer can be used.

“Each different setting has different materials, and and all of that has gone into the systems learning process,” Wicks said. “They’re thinking about application on both small and larger and larger and larger levels, so eventually it will become a stationary system that matches the specific needs of the site that it is situated at.”

While rStream is still in its early phases, Goodine sees this technology replacing consumers’ need to know what goes into the recycling bin.

“I think there’s an opportunity for this to be everywhere there’s garbage,” Godine said.