Waste bot, want bot
Using artificial intelligence to sort recycling might hold the key to a circular economy. Will Simpson presents some of the first case studies from the field.
The robotic age in recycling has always seemed to be just around the corner. There have long been predictions that artificial intelligence will replace human labour along the picking and sorting lines (and in some other areas too). At present, though, most UK MRFs still get by on a combination of automated sorting – near infra red, eddy current separators etc – and human labour.
But a confluence of events looks like it might finally force full automation upon the sector. In recent years, many countries – China being the most notable – have set tougher standards in terms of the purity of imported recyclables. Combine this with the increasing cost of residual waste disposal due to the landfill tax and a post-Covid, post-Brexit labour shortage that has driven up labour costs and you have makings of a crisis.
Then there’s the AI itself, which, especially in regard to camera technology, has made technical advances over the last decade. Extended Producer Responsibility is also having a knock on effect in that recyclate purchasers are demanding better quality and more accurate figures in terms of data regarding contamination and purity levels. The days of relying on guesswork and estimates are coming to an end. If the recycling market is to function properly – and by extension any hope of establishing a truly circular economy – then it looks increasingly dependent on the adoption of new technology.
Victor Dewulf, the CEO of London-based AI and robotics startup Recycleye, is fully aware of this: “If I sell you a barrel of oil today you’ll know what you’re buying. I know what I’m selling. All too often, though, if there is a tonne of plastics you don’t know exactly what you’re getting. But people like Veolia, people like Biffa, need to know what they’re getting.”
With an accuracy rate of around 97 per cent, the robot Recycleye has developed is, he claims, the solution to this. It can also process up to 60 images per second and recognise up to 28 ‘waste material classes’. Plus it’s more lightweight than other competitors – almost 75 per cent less, according to Dewulf – and its vision system is able to sit ‘on top’ of an existing conveyor, thereby cutting down on capital cost.
As you’d expect, the Recycleye robot incorporates machine learning into its system: “When it sees a mug and it knows it’s a mug, it will tell us the degree of certainty with which it thinks that. So if it’s 99 per cent sure, cool. If it’s only five per cent sure then the image goes into our database and somebody will check it. There is always a human element there as back up. There is a myth, I think, that AI learns by itself. It doesn’t, it only learns if it has better data.”
Recycleye currently has two robot systems in operation at UK facilities. The first to open was at re3 in Berkshire in September. This is a waste management partnership between Bracknell Forest, Reading and Wokingham Borough Councils and FCC Environment. Re3 expect to have their Recycleye system fully up and running before the end of 2021.
“Currently we’re looking at it sorting plastics,” Rory O’Brien, re3’s General Manager explains. “Though it can be utilised to sort a wide range of materials from plastic bottles, tubs and trays, foil trays, aluminium, cardboard, paper, aerosols and steel cans. The AI system sits on top of one of the facility’s existing conveyor belt, providing the facility with total visibility on the waste streams contained on that belt.”
“The robot not only has this ability to pick specific grades of materials, it can also pick brand types if we are required to do so in the future. This will bring clear advantages and future proofing as it allows us to identify and separate out additional materials, should this be a requirement, in particular when you consider the potential changes that may be delivered by the Resources and Waste Strategy and Environment Bill.”
The Recycleye robot, though, is far from the only one on the market. There is also a system developed by the UK firm Greyparrot, which can identify up to 50 different ‘waste categories’. Rather than actual sorting, the Greyparrot system simply provides data concerning composition along a real time flow, including product, material and brand level. It claims to have a 95 per cent success rate.
Then there is AMP Robotics, a US firm that has so far concentrated on the North American market but is looking to expand into Europe. Their system is broadly similar to Recycleye’s in that it can recognise objects based on shape, waste stream type and brand: “It can do exactly what a human can do,” claims Gary Ashburner, AMP’s European General Manager. “It’s able to look at something that is clearly cylindrical, red, reflective and with a black hole in one end. It might see some words – ‘Co’ – and pick out that it’s a Coca-Cola can. It can pick out other interesting distinguishing characteristics like names, bar codes and colours just like any human would. There’s no material type that AI struggles with or is good at implicitly. It’s all about the training it recieves.” This is machine learning, a crucial part of AI – AMP have a team of annotators that can feed any new information into the robot’s ‘brain’.
Any AI network is built by recognition of items. The larger your network, the larger your deployed cameras, the better your AI is as the robot is seeing more images and the computer is learning from them all the time. The fewer images it sees, the less it learns and the lower its accuracy.
Going hand in hand with this has been improvements in camera technology. “The huge difference now compared with the cameras of 10, 20 years ago is the quality of images you’re looking at,” says Ashburner. “That and the speed and the sharpness of the image.”
He suggests that a single AMP robot can replace two human pickers, with a dual or tandem model replacing four. “Labour is obviously a real concern at the moment,” he says. “You can’t get people to do the job or if you can they soon move on. It’s a dirty job so you’ve got hygiene issues – pickers can pass on contamination. And you’ve got the health and safety issues – you’ve also got the risk of injury from sharps, razor blades and even needles.”
“So, as well as eliminating labour costs, you’re eliminating the health and safety issues. You’re picking more material so you’re either picking higher value material faster or eliminating the contamination on the bales. And when you are putting it on the market your bale has got a better purity level so you’re getting more money for it.”
Ashburner claims that the AMP system has around a 90 per cent accuracy rate, but the big differentiator, he claims, is its speed. “That’s where we are years ahead of everybody else. We can recognise material – and more diverse types of material – quicker, and the robots can be accurately positioned to pick that material.”
The only downside of AI and robotics in general – apart from the obvious loss of human jobs – is the capital cost. Inevitably, installing a robotics system is a huge outlay, one that in some cases may necessitate a significant re-design of the whole MRF. Both Dewulf and Ashburner claim that their respective systems have around a three year return on investment if a facility is able to convert to running on a 24/7 basis. “If you’re only looking at two shifts five days a week then you’re talking maybe five years,” says Ashburner.
The first AMP robot in the UK was installed at Recyco, a Northern Irish facility, in September 2021. Recyco’s CEO, Michael Cunningham has reported that the AMP system has doubled Recyco’s pick rate overnight and that he is “blown away with the identification accuracy of the AI.”
Ashburner relates that AMP is “getting a lot of interest from the UK at the moment.” With an increased demand for accurate data, the recycling sector here appears to be on the cusp of significant step change. Whilst the robots aren’t going to take all our jobs overnight, it looks likely that they will soon consign manual picking and sorting lines to history. If that results in purer, better quality recyclates then it surely has to be a price worth paying.