Introduction to AI in recycling
AI in recycling is addressing a practical problem: identifying useful material quickly enough to recover it from a moving waste stream. Computer vision can distinguish objects, assess their characteristics and guide sorting equipment. It can also give facility operators a more detailed account of what passes through their plants. Those functions connect recognition to two different outcomes: physically recovering materials and improving decisions about the recycling process.
A November 10, 2026 NVIDIA report describes four startups applying these methods to plastics, household recyclables and used clothing. AIRECO sorts plastic for potential reuse in the same kind of product. EverestLabs AI guides robots on recycling lines. Greyparrot measures the contents of waste streams. reverse.fashion evaluates garments to determine their potential uses and value.
Together, these examples show how AI can support waste management at several stages. Their evidence also needs careful interpretation. All four companies belong to NVIDIA’s Inception startup program, and the account comes from their computing supplier. Reported results include company estimates, customer-reported gains and pilot measurements, alongside targets for future production. They do not constitute an independent comparison of the systems.
The central question is therefore more specific than whether AI can recognize rubbish. It is whether recognition helps a facility recover more useful material, sort it into a more valuable category or make better operational decisions. The four applications offer different answers, with different measures of success.
Impact of AI on waste sorting
Sorting determines which materials can remain useful and in what form. NVIDIA’s account describes inconsistent incoming plastic waste, difficulty finding manual sorting labor and mixed scrap that becomes lower-grade products. AIRECO’s approach is to separate plastics more finely so that recovered material can enter closed-loop recycling: reprocessing a used item into a new version of the same product.
This distinction matters because recovering an object does not establish what it can become. A bottle returned to bottle production follows a different path from plastic converted into a product that is harder to recycle again. Finer classification is intended to make the former route more achievable. AIRECO’s founders argue that better sorting can improve quality and reduce cost, making recycled plastic more commercially useful.
The operational challenge is timing. At material recovery facilities, which separate household recyclables, items move along conveyor belts while the system identifies them and sorting equipment acts. EverestLabs AI’s deployment illustrates this connection between computer vision and robotics: a correct classification must arrive while the object remains available for picking. Recognition and physical recovery are related tasks, but they are not interchangeable measurements.
Greyparrot demonstrates a second role for vision AI. Its Analyzer platform identifies material and brand from above the conveyor, producing information about the waste stream. NVIDIA contrasts this with operators manually sampling less than 1% of the material. More extensive observation gives operators a basis for decisions that a small sample may miss, particularly as the contents change by supplier, season or hour.
Textiles introduce another distinction. Identifying a garment’s general type is insufficient when its value depends on small defects, construction and style. reverse.fashion uses high-resolution vision to assess these attributes and assign garments to categories. Here, the sorting decision concerns potential value and use as well as material identity.
Across these examples, recycling efficiency has several meanings: classification accuracy, sorting speed, material yield and operational performance. A high score on one measure should not be read as proof of all the others. The reported deployments are most informative when their results stay attached to the particular task and conditions being measured.
Case studies of AI applications in recycling
AIRECO: finer plastic separation on a pilot line
Taiwan-based AIRECO uses vision AI at a recycling plant in Tainan. NVIDIA Jetson Orin NX modules run inference on its pilot line, including supporting robotic arms that remove selected objects from a conveyor. NVIDIA reports recognition accuracy above 95% on that test line.
That figure describes recognition under the reported test conditions. The account does not supply a detailed accuracy methodology or establish the proportion of all incoming material successfully recovered. It supports a claim about the pilot’s recognition performance, rather than a complete measure of plant economics or recycling outcomes.
AIRECO’s expansion plans are more ambitious. For a full production line planned for 2027, the company aims to raise annual waste processing capacity from 18,000 to 48,000 tons and cut operating costs by 85%. These are targets, not results already achieved. Their significance is that AIRECO is linking finer sorting to both production capacity and commercial viability. Whether the planned line delivers those gains remains unresolved in the report.
EverestLabs AI: robotics and operational information
U.S. startup EverestLabs AI deploys computer vision-guided robots on material recovery lines. NVIDIA reports about 150 robots installed across North America and Australia. The company runs inference on NVIDIA RTX GPUs at the edge, while training uses NVIDIA data center GPUs and optimization with NVIDIA CUDA.
This separates the work of training models from their use on the sorting line. The real-time requirement comes from the moving conveyor: the system must identify an object and support a pick before it passes. EverestLabs says it continues improving models to classify and pick different objects while recording information about material type, brand and packaging.
The company estimates that its deployed systems have identified about 40 billion objects, recycled roughly 25,000 tons of material and avoided about 30,000 metric tons of carbon dioxide emissions. These are cumulative company estimates. NVIDIA’s account does not provide the methodology needed to independently assess the emissions figure or compare it with another deployment.
EverestLabs is also shipping agentic AI tools intended to turn operational data into actionable information for floor managers and executives overseeing multiple facilities. The report establishes that direction but gives little detail about the tools’ decision-making authority or specific workflows. It supports describing an assistance layer for staff, without assuming autonomous control of a plant.
Greyparrot: measuring the material stream
U.K. startup Greyparrot mounts Analyzer above conveyor belts to identify passing objects by material and brand. According to NVIDIA, it operates in dusty conditions or poor lighting on belts moving at up to 4 meters a second. NVIDIA GPUs support model training on billions of objects, and NVIDIA Jetson modules run detection at the edge.
The reported installed base exceeds 300 Analyzer systems in more than 20 countries. Those systems have made over a trillion bounding box detections. This is a measure of detection activity; it should not be substituted for a count of objects physically recovered or tons recycled.
NVIDIA reports that Analyzer reduces the time required to check 15 tons of waste from 375 hours to six. Facilities using its real-time data report efficiency gains of 10% to 30%, while packaging company Amcor reportedly increased its yield of food-grade plastics by 12%. The account does not give a common definition or measurement protocol for the facility efficiency figures, so they should remain attributed outcomes rather than a universal expectation.
Greyparrot’s users also include Coca-Cola, Colgate-Palmolive and Kenvue, which use its technology to measure how their packaging is recycled in sorting plants. This extends the information’s audience beyond plant operations to companies seeking evidence about packaging outcomes.
reverse.fashion: evaluating garments at high resolution
Berlin-based reverse.fashion applies computer vision to used clothing. Its system identifies product type, defects, color, size, brand, fabric construction, season and style, then assigns each garment to its most valuable category. The task demands attention to small details because a defect can affect a garment’s quality and potential value.
Each textile passes through more than a dozen vision AI models as it moves along a conveyor at 3 meters a second. Those models run on NVIDIA Jetson modules at the edge and are optimized with NVIDIA TensorRT. Training takes place on NVIDIA GPUs in workstations and the cloud.
Cofounder and chief technology officer Karsten Pufahl estimates that the system sorts in an hour what previously required a week. NVIDIA says it is rolling out to facilities in Europe. The speed comparison is his estimate, and the account does not specify the quantity of textiles or working-hour assumptions behind it. It indicates a claimed improvement in sorting productivity, not a standardized benchmark against the other systems.
Future of AI in waste management
These deployments point toward a closer relationship between material recognition, sorting equipment and operational analysis. This is an interpretation of the applications described, rather than evidence that every recycling facility will adopt the same architecture. AIRECO and EverestLabs connect recognition to robotic picking; Greyparrot emphasizes measurement; reverse.fashion uses detailed classification to support garment valuation.
Their computing arrangements share a recurring pattern. Models are trained using GPU resources, while inference runs near the conveyor on edge hardware. The specific products vary: AIRECO uses Jetson Orin NX, EverestLabs uses RTX GPUs for inference, and Greyparrot and reverse.fashion use Jetson modules. The account documents those choices but does not provide comparative hardware benchmarks or establish that one setup is best for every waste stream.
The agentic element is narrower than the broader use of AI. EverestLabs explicitly describes tools for providing actionable data to staff. The other examples primarily document computer vision, classification and robotics. Calling all of them autonomous agents would obscure what the evidence actually shows. A useful next question is how operational information influences decisions and who retains responsibility for acting on it; the report does not answer that in detail.
For an operator assessing these systems, the cases suggest matching evaluation to the intended outcome. A plastic sorting project needs evidence about recovered material and its suitability for reuse. A monitoring system needs evidence that its information improves decisions. A garment classification system needs to demonstrate useful categorization as well as speed. These are evaluation implications of the documented tasks, rather than additional performance claims.
Several questions remain open. AIRECO’s production targets still require delivery. Reported gains across the companies use different metrics and conditions. The evidence does not supply a consistent account of purchase costs, maintenance requirements, payback periods or model errors. It therefore supports examining specific applications, but not a general return-on-investment promise for AI in recycling.
Conclusion on sustainability and AI
The clearest contribution of these systems is making waste streams more legible and, in some applications, connecting that understanding directly to sorting. Plastics can be separated more finely, robots can target household recyclables, operators can inspect a broader stream of material data, and garments can be assessed for uses that depend on their condition and characteristics.
The sustainability case depends on what follows recognition: useful material recovered, suitable feedstock produced or garments directed toward valuable uses. NVIDIA’s report offers concrete examples and attributed results along that path, alongside expansion targets and unresolved measurement questions. It does not establish that AI alone guarantees circularity.
For waste management, the practical promise is a better-informed sorting process. Assessing that promise requires keeping pilot accuracy, company estimates, customer outcomes and future plans distinct—and judging each system by the material outcome it is meant to improve.
