Summary
NVIDIA profiles four startups using computer vision and robotics to sort plastics, household recyclables and used clothing. Their reported results range from faster material checks to higher sorting capacity and improved recovery of specific materials.
NVIDIA has profiled four startups applying computer vision and robotics to recycling and textile recovery, from identifying plastic on a conveyor belt to grading used garments for resale or reuse. The examples span pilot and operating systems, and their reported results measure different things: recognition accuracy, sorting speed, material yield and estimated volumes processed.
From plastic picking to garment grading
Taiwan-based AIRECO uses vision AI on a pilot line at a recycling plant in Tainan. NVIDIA says the system runs on Jetson Orin NX modules and guides robotic arms that pick target plastics from a conveyor. AIRECO reports recognition accuracy above 95% on the test line. The company’s planned full production line aims to increase annual processing capacity from 18,000 to 48,000 tons and cut operating costs by 85%; those are targets, not reported results from an operating production line.
U.S.-based EverestLabs AI uses computer vision-guided robots to sort items at material recovery facilities. The company has about 150 robots deployed across facilities in North America and Australia. It runs AI inference on NVIDIA RTX GPUs at the edge and trains models on NVIDIA data-centre GPUs. EverestLabs 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.
Greyparrot’s Analyzer takes a different approach: cameras mounted above conveyor belts identify passing objects by material and brand, giving operators a continuous view of what is moving through a facility. The company says more than 300 systems in over 20 countries have made more than a trillion bounding-box detections. It reports that checking 15 tons of waste can take six hours with Analyzer, compared with 375 hours previously. Facilities using its real-time data report efficiency gains of 10% to 30%; at packaging company Amcor, the system helped increase the yield of food-grade plastics by 12%, according to NVIDIA’s account.
Berlin-based reverse.fashion applies computer vision to used clothing. Its system assesses characteristics including garment type, defects, colour, size, brand and fabric construction, then assigns items to categories with different potential uses. NVIDIA says the models run on Jetson modules and are optimized with TensorRT software. The startup estimates that its system can sort in an hour what previously took a week; it is rolling out to facilities in Europe.
Why sorting detail matters
These systems address different steps in recovery. Computer vision can classify material or product characteristics, while robots can act on those classifications by picking items from a moving belt. For textiles, detecting small defects and features can help route garments to resale, reuse or lower-value processing. Running inference at the facility—often called edge computing—means the system analyzes items locally as they pass, rather than relying on a remote service for each decision.
The examples suggest how more detailed, continuous sorting could help facilities separate materials that manual sampling or broad categories may miss. But the reported measures are not directly comparable: they cover different materials, tasks and outcomes, and the systems range from a pilot line to deployments across multiple facilities. The capacity and cost figures AIRECO aims for will depend on its planned production line.