Raspberry Pi fans have never been short of ideas to put the tool to good use, for applications as wacky as they are useful. Now researchers are fitting the low-cost computer with artificial intelligence, high-resolution cameras and robots, to sort through rubbish and reduce the amount of waste going to landfills.
Engineers from Liverpool Hope University played around with a Raspberry Pi 3 model, combining the device with optical sensors and computer vision algorithms, to create a tool that can distinguish between paper, glass, plastic, metal and cardboard.
Set up in a material recovery facility (MRF), where household rubbish is usually sent to be sorted, the technology could spot different materials on the conveyor belt where waste is dumped, and accordingly instruct robots to recycle specific objects as they come towards them.
Karl Myers, from Liverpool Hope University’s department of mathematics and computer science, told ZDNet: “It is designed to be integrated with any of the robotic systems that are on the market at the moment. The Raspberry Pi sends a signal via serial communication to the robotic arm about the position of the recyclables, and the robot just grabs the object.”
The researchers said that the algorithm achieved up to 92% success rate, with a baseline performance of 90%, and argued that the approach was therefore viable for commercial use.
With the growing amounts of rubbish that are generated every year across the world, improving the performance of recycling facilities is key to avoid disposing of waste in landfill sites. Of the 229.9 million tons of solid waste generated in the UK in 2017 alone, only 47% were recycled. The Department for the Environment has set a goal of pushing recycling rates to 50% for 2020.
Currently, households in the UK are asked to keep all recyclables – paper, metals, glass and so on – in a single receptacle, which is collected from doorsteps before being sent to a MRF for sorting and processing.
Once there, the materials are placed on a conveyor belt, where cardboard, containers, paper and plastics are removed manually. Tin and steel cans are then sucked off by a powerful magnet, while a reverse magnet causes aluminum cans to fly off the conveyor and into a storage container.
Human operators supervise the process, which often comes with errors, inefficiencies and extra costs. “The longer-term vision would be to remove humans entirely from MRFs,” said Myers. “Only about 40% of the recycling that we send to MRFs is actually recycled, and this is because of human interactions. Our system will improve the accuracy in the MRF because it removes the inadequacies of the human.”
Myers and his team trained the algorithm with a database of 3,500 different images of rubbish, combining a resource called TrashNet with images from Google. The researchers used transfer learning, a particular approach in machine learning that enables the AI system to store the knowledge gained solving one problem, and apply it to solve a new, different (but related) one.
The method mimics the human brain: if you are learning how to use a motorcycle, the chances are you will call on your knowledge of riding a bike. With transfer learning, the algorithm can similarly use knowledge gained from a previous problem to solve a new one that has little data.
“It removes the individual learning paradigm,” said Myers. “In this case, it means that no training whatsoever is required for the system – it will use all the images and past knowledge from other datasets and apply it to the problem it is working on. It’s essentially a plug and play.”
Useful, accurate, and easy to deploy, therefore – but also cheap. Building up the system cost less than £100 in total, which the researchers believe will largely boost adoption of the technology across the world.
There is one draw-back: based on the engineers’ tests so far, Raspberry Pi-controlled recycling robots will be slower at sorting through rubbish than humans are. Myers argued, however, that a slow technology is still “perfectly adequate”, so long as it is accurate. According to the researcher, the efficiency and cost of the technology offset the system’s poor speed performance.
The research team is now hoping to build up their invention to integrate it with waste retrieval solutions such as satellite-controlled platform SeaVax, which is designed to roam the oceans and operate like a giant vacuum cleaner.
SeaVax doesn’t fetch objects intelligently; rather, the robot indiscriminately scoops up the objects it finds and throws them into its hopper. Fitting the platform with the automated recycling technology developed by Myers and his team could add great value to the system, by sorting through waste directly at the collection point rather than in an MRF.
Some engineering challenges remain before the researchers can achieve this goal, but Myers is already testing extending the invention and testing the algorithm on more powerful systems. In the meantime, the engineer is confident that the technology will be met with positive feedback, even in its current state.
“If this goes into adoption, it will remove substantial costs associated with manpower,” he says. “With the cost of it being so low I can only imagine the industry will pick it up.”