EverestLabs is using physical AI to tackle one of the world’s most critical problems: recycling.
The company deploys robotic arms on the sorting lines of recycling plants, pulling aluminium, plastics and contaminants from fast-moving waste streams. With close to 250 systems in the market, its robots are capable of up to 70 picks per minute with a 95% success rate.
“We’re not just building technology for technology’s sake,” EverestLabs CEO R. Paul Singh tells IZON in an interview.
“This is ‘useful AI’. We have logged more than 10 million belt hours and recovered more than 100 billion materials which has contributed to a more sustainable and circular economy.”
EverestLabs was founded in 2018 in California and some of its systems have been in production for more than five years across the US.
“I don’t think anybody else has that history of real-world deployments in this industry,” Paul says.
Hitting the scaling curve
Paul is an entrepreneur “in the valley” and was appointed CEO of EverestLabs in December 2025. He’s had four successful exits as a founder, with one going public on NASDAQ, and so took over from the previous CEO Jagadeesh ‘JD’ Ambati to lead the next phase of growth.
Paul says: “Scaling a company is an exciting journey, and experience navigating growth across different disciplines can help build on its momentum. That’s the perspective I brought, and I believe it resonated with the board and founder as they looked ahead to the company’s next chapter.”
Paul describes EverestLabs as a software company which is vendor agnostic. “Robotics is a means to an end for us. We can drive any robot arm to do what we want it to do, and we had to test and see which one was the most industrial that could last in these harsh environments.”
“It’s not an air conditioned room”
These Materials Recovery Facilities or MRFs are notoriously chaotic environments, featuring rapidly moving conveyor belts that are flooded with unpredictable, contaminated waste streams.
“It’s not an air-conditioned room by any means and it’s not an environment where human beings should really work and most of them don’t want to work there. Sorting recyclable materials is demanding, hands-on work. Automation can help create a cleaner, safer workplace while supporting around-the-clock operations.”
How the robots work
EverestLabs’ robotic arms use 3D depth-sensing cameras and AI vision software to identify, track and pick recyclable materials directly from the moving conveyor belt.
Once a target item is identified, the system directs the modular robotic arm to suction-lift the specific material off the belt and drop it into the proper chute or bin.
“A human sorter averages about 20 to 25 picks a minute,” Paul says. “Our robots complement that work by delivering consistent performance throughout the shift, averaging about 45 picks a minute and reaching 60 to 70 under favourable conditions. Pick rates depend on the type of material and how it is presented on the conveyor.”
Each plant will have different use cases. The Last Chance Recovery Line picks out recyclables just before material is sent to landfill; a quality control line is placed at the end of the sorting process to inspect and protect final bale quality; container lines automate the bulk separation of plastic, metal and glass; and purity boosting is designed to target highly strict material streams.
Mattresses, shoes and hoses
The variety of items arriving at a facility adds another challenge. “Mattresses are very common, believe it or not,” says Paul. “Shoes, hoses.” These items can enter recycling streams even though they require different collection or disposal methods, potentially clogging equipment and interrupting operations. Batteries pose a particular concern because they can cause fires.
Material presentation, visibility and safety are essential to efficient recycling operations. Paul says the vision system achieves 95 to 96% accuracy, with pick success rates of 80 to 95%, depending on operating conditions.
“Lighting and material depth are key factors,” he explains. “Cameras can only see the exposed layer.” EverestLabs provides dedicated lighting, while upstream equipment can spread materials to improve visibility and access.
Never selling against the human
The robots may outpace people on the line, but EverestLabs does not pitch them as a like-for-like replacement for a worker.
“We never sell against the human, to be very frank,” Paul says. “We’re solving the issue of not having enough labour wanting to work in these environments, more than what the cost is.”
The economics still have to stack up, though. A plant paying around €50,000 a year per sorter needs a robot that costs no more. “If your plant only runs four hours a day, it’ll be hard to justify robotics in most cases. If you run two or more shifts, it’s very easy to justify.”
Once the robot is installed, EverestLabs monitors every unit around the clock, with a human kept in the loop, so it can show customers the pick rate they were promised.
Teaching the plant to think
Customers are starting to see themselves less as waste managers and more as materials producers, Paul says, and that means they want data. This was the thinking behind Navigator, which EverestLabs launched in August as what it calls the first multi-agent AI platform built for materials processing and recycling facilities.
Navigator combines vision models at the belt, vision-language models that interpret each material stream, and conversational agents that recommend actions and can carry them out, if needed. It is designed to integrate with existing plant control networks and older equipment, helping limit the need for retrofits and production interruptions.
EverestLabs claims it can lift a facility’s throughput by 20 to 30%, predict belt and equipment failures before they happen, and pinpoint where valuable material is being lost to landfill.
The next 12 months
Over the next year, Paul says the team is focused on expanding the number of partnerships in the US and pushing into Europe, where he hopes to become the market leader.
Further out, he sees the physical AI industry splitting between generalist companies selling a “robot brain” and end-to-end system players like EverestLabs that can own one industry. A general-purpose robot placed on a sorting line, he says, “will fail in most cases”.
“Robots for specialisation versus general-purpose robots is where the battle will be fought, and I think special-purpose robots will win in most cases.”
