Trash facilities are using edge cameras and robotic arms to pluck value out of fast belts of waste. In a November 10 update timed with COP31, NVIDIA shared how four Inception startups sort plastics, food packaging, and old clothes at plant speeds.
Global waste tops 2 billion tons per year, yet less than 7% of raw materials in the economy come from reused stock. Sorting belts move too fast for manual checks, leaving clean plastic and worn fabric buried in mixed trash.
Key Takeaways: AI Waste and Textile Sorting
- AIRECO (Taiwan): Pilot lines hit over 95% accuracy on NVIDIA Jetson Orin NX chips, with plans to grow plant throughput from 18,000 to 48,000 tons next year.
- EverestLabs AI (US): Runs roughly 150 robots across North America and Australia and estimates 40 billion items spotted, 25,000 tons recycled, and 30,000 metric tons of CO2e saved.
- Greyparrot (UK): Deploys 300+ Analyzer camera units across 20+ nations, cutting 15-ton waste audit times from 375 hours down to 6 hours.
- reverse.fashion (Berlin): Uses Jetson modules and over a dozen TensorRT models to sort used clothes at 3 meters per second, finishing in one hour what used to take a week.
How the Four Sorting Systems Compare
Standard plants sample less than 1% of trash by hand because belts move up to 4 meters per second in heavy dust. Edge vision chips fix that blind spot by running live checks right over the line.
| Startup | Target Stream & Edge Hardware | Self-Reported Impact |
|---|---|---|
| AIRECO | Mixed plastics on Jetson Orin NX | >95% pilot accuracy; aims for 85% cost drop |
| EverestLabs AI | 150 MRF robots on edge RTX GPUs | 40B items read, 25k tons saved, 30k tCO2e cut |
| Greyparrot | 300+ camera units on Jetson | 10% to 30% plant gains; +12% yield at Amcor |
| reverse.fashion | Used clothes via TensorRT + Jetson | 3 m/s speed; 1 week of work done in 1 hour |
It helps to tell firm sales claims apart from outside plant data. Each startup shared its own counts through NVIDIA. Outside trade notes on Waste Dive confirm that smart picking arms help on busy final belts where tired crews miss clean scrap.
Fine Plastic Sorting and Plant Data
Old plastic can only beat new oil plastic when sorted loads stay clean. Taiwan firm AIRECO runs smart arms on a test line in Tainan built on Jetson Orin NX boards. CEO Chun-Feng Liu said clean splits turn bottles back into bottles instead of low-value scrap. His team aims to cut run costs by 85% once a 48,000-ton full line opens next year.
Across North America and Australia, EverestLabs AI runs about 150 picking arms on RTX edge GPUs. CEO R. Paul Singh noted that fast belts do not wait for slow arms. That pace makes plant sorting a tight real-time race. His firm estimates its units have read 40 billion items and saved 25,000 tons of scrap so far.
UK firm Greyparrot skips arms and sets smart cameras right over mixed trash belts. Over 300 Analyzer boxes across 20 nations have tracked a trillion items on belts moving up to 4 meters per second. Food packer Amcor saw a 12% yield boost in clean plastic using live belt charts. That tool also cut 15-ton load audit times from 375 hours to 6 hours.
Putting smart chips in dirty plant zones tracks a wider industry trend. Read our stories on 3,500 AI laser robots on salmon farms and NVIDIA’s South Korea plant AI push.
Automating Europe’s Textile Pileup
Europe discards nearly 7 million tons of old clothes every year. Only about a quarter is collected apart and graded by hand into hundreds of resale and cloth grades.
Berlin startup reverse.fashion built a high-res camera belt that scans clothes moving at 3 meters per second. Each item passes through over a dozen vision models tuned with NVIDIA TensorRT and Jetson hardware.
CTO Karsten Pufahl noted that clothes need sharper scans than scrap bottles because buyers reject small snags or stains. His system inspects brand, fabric weave, color, size, and flaws to route vintage gear to top buyers. It finishes in one hour what hand sorters needed a week to grade.
Modeling physical materials at speed also reaches research labs, as seen in our coverage of MIT’s WorldParticle matter simulator. At the same time, rising data center compute keeps putting strain on power grids, a trade-off explored in our look at Cornell’s study on AI power grid costs.
Frequently Asked Questions
Which startups did NVIDIA highlight for AI recycling?
NVIDIA featured four Inception members: Taiwan’s AIRECO, US robotics firm EverestLabs AI, UK waste data leader Greyparrot, and Berlin textile sorter reverse.fashion.
How fast can edge vision systems inspect passing trash?
Greyparrot cameras track objects on conveyor belts moving up to 4 meters per second, while reverse.fashion grades passing garments at 3 meters per second.
Are the recovery totals verified by independent auditors?
Figures like 40 billion items for EverestLabs and 85% cost cuts planned by AIRECO come from the startups and NVIDIA, though plant partners like Amcor backed the 12% food-grade plastic gain.
What NVIDIA hardware runs inside these recycling plants?
Startups train object models on data center GPUs and run live inference at the plant edge using NVIDIA Jetson modules, Jetson Orin NX boards, and RTX GPUs with TensorRT.
