Greyparrot Tracks Waste Streams to Recover $120 Billion in Packaging

Uncovering Waste with AI: The Climate Crisis—And Opportunity—No One’s Talking About With Greyparrot — Ai4

Ai4

Closing the post-disposal data gap with conveyor-belt computer vision is essential to automate sorting facilities and recover $120 billion in discarded commodities that generate 10% of global greenhouse emissions.

AI-assisted summary of Uncovering Waste with AI: The Climate Crisis—And Opportunity—No One’s Talking About With Greyparrot — original source. Check the original for full context.

Source published: 2025-08-28

Detailed notes

## Background Published on YouTube by Ai4 on August 28, 2025, this keynote features Greyparrot co-founder Ambarish Mitra discussing how artificial intelligence can solve the opaque global waste crisis. The description highlights how digitizing waste flows across recycling facilities provides real-time data on discarded materials to accelerate a circular economy.

## TL;DR Ambarish Mitra breaks down the massive environmental and economic costs of municipal waste, where discarded materials generate nearly 10% of global greenhouse gas emissions while billions of dollars in valuable resources are lost. He details how Greyparrot built a specialized Large-Scale Computer Vision Model trained on over 100 million images to identify crushed, soiled objects across 111 material categories on high-speed conveyor belts. By acting as an intelligent operating system for Material Recovery Facilities and providing packaging visibility through the DeepNest platform, the technology automates sorting facilities and creates empirical accountability for consumer brands and policymakers.

## Atmosphere & style The presentation carries an urgent, clear-eyed, and pragmatic tone that replaces vague sustainability platitudes with rigorous data and industrial realities. Mitra speaks with an authoritative insider perspective, blending Silicon Valley technological ambition with the gritty, harsh operating conditions of recycling conveyor belts. The narrative arc moves from confronting staggering systemic failures to offering a compelling, solutions-driven vision of how real-time intelligence can transform waste from an invisible liability into a circular asset.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) Post-consumer municipal waste is a massive, overlooked driver of the climate crisis, generating nearly 10% of global greenhouse gas emissions—more than double commercial aviation. - [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) Over 9 million people die annually from waste-crisis fallout, exceeding the combined global death toll of AIDS, tuberculosis, and malaria. - [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) Between $80B and $120B in economically viable materials are lost globally every year due to inadequate sorting and processing infrastructure. - [04:07](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=247s) A massive data asymmetry exists between the $23T consumption economy, which tracks consumer actions by the millisecond, and the post-bin 'shadow economy' where data tracking vanishes. - [04:07](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=247s) The term 'plastic' encompasses 67 to 68 common chemical variants that do not recover equally, often proliferated for shelf marketing rather than recyclability. - [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) Standard AI vision models fail on waste streams because crushed, soiled, and distorted objects—such as a soda can with over 20,000 crushed variations—do not exist in standard web training sets. - [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) Greyparrot built a specialized Large-Scale Computer Vision Model (LSCVM) trained on over 100M images to recognize 99.9% of manufactured items across 111 material and grade categories. - [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) Material Recovery Facilities (MRFs) operate under harsh conditions with conveyor belts moving at 3 m/s in extreme heat and dust, requiring low-cost, real-time AI to automate optical sorters and robotics. - [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) Historical recycling reports and UN assessments have often relied on statistically flawed sample sizes as small as 500 items to draw macro-level conclusions. - [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) Cosmetic packaging design choices heavily impact recyclability; for example, Sprite's switch from green to clear PET bottles increased material recovery rates by 40%. - [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) Granular, empirical waste intelligence enables verifiable purity thresholds for secondary commodity markets, accurate carbon accounting, and evidence-based environmental policy.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) The Overlooked Crisis of Post-Consumer Waste

Ambarish Mitra introduces his background as co-founder of Greyparrot, recalling how he built the world's first visual search engine in 2012 to map physical objects in their pristine states. By 2019, he realized that an entire undocumented world existed that nobody was discussing: the reality of post-consumer municipal waste. While the mainstream cultural imagery of climate change fixates on oil rigs, smoking industrial chimneys, and overt pollution, everyday trash thrown into household bins is routinely overlooked.

Every human, across all cultures and geographies, leaves an unavoidable physical footprint through daily consumption. Globalization ensures that virtually every basic good arrives wrapped in some form of packaging. Consumers often believe their environmental responsibility concludes the moment an item is dropped into a bin, but that disposal is actually only the beginning of an opaque, poorly understood journey that carries severe environmental and human consequences.

- Over 9 million people die annually from waste crisis-related fallout, surpassing the combined global death toll of AIDS, tuberculosis, and malaria. - Between $80 billion and $120 billion worth of economically viable materials are discarded globally each year due to inadequate processing infrastructure. - In the United States alone, roughly $9.5 billion in valuable material packaging is lost annually from retail, DoorDash, Amazon, and other consumer deliveries. - Decomposing or incinerated landfill waste generates nearly 10% of global greenhouse gas emissions—more than double the emissions generated by the entire commercial aviation sector.

---

[04:07](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=247s) Policy Realities and the $23 Trillion Consumption Gap

Arriving directly from the sixth round of negotiations for the UN Plastics Treaty in Geneva, Mitra highlights the urgent push to hold manufacturers accountable through plastic reduction mandates, recoverability standards, and producer taxes. The generic term "plastic" masks an intricate chemical ecosystem of 67 to 68 common variants—including PET, HDPE, polypropylene, and polystyrene. These materials do not recover equally, and the proliferation of resin types, dyes, and sleeves has historically been driven by corporate branding and shelf marketing rather than recyclability.

```mermaid flowchart TD accTitle: The Consumption Economy vs The Shadow Economy subgraph Consumption["Consumption Economy ($23T Value)"] A["Consumer Purchase & Tracking"] --> B["High-Velocity Ad & Market Data"] end subgraph Shadow["Shadow Economy (Unmeasured Waste)"] B --> C["Disposal in Waste Bin"] C --> D["Treated as Zero/Negative Asset"] D --> E["Material Recovery Facility (MRF)"] E --> F{"Automated Sorting & AI"} F --> G["Circular Economy Stream"] F --> H["Landfill / Incineration"] end ```

The fundamental bottleneck lies in the asymmetry between the consumption economy and what Mitra calls the "shadow economy." In the $23 trillion global consumption economy, consumer actions are tracked down to the millisecond by dozens of corporations targeting future purchasing intent. Yet the instant an item enters a bin, data collection vanishes. The object shifts from an asset with an explicit purchase price to a zero-value or negative-value liability that municipalities must pay to discard. Greyparrot was founded to close this tracking gap, deploying continuous monitoring to trace everything from a chocolate wrapper to a 100-inch television box across global metropolitan waste streams.

---

[08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) Building a Large-Scale Computer Vision Model for Destroyed Objects

Managing waste is uniquely challenging because no single global industry interfaces directly with 7 billion individuals producing a bag and a half of trash every week. Because traditional human labor and mechanical systems cannot handle this volume in real time, waste analysis was ripe for artificial intelligence. Standard Large Multimodal Models and vision APIs fail completely when confronted with waste imagery because crushed, soiled, and torn objects do not exist in standard web training sets. A standard Coca-Cola can, for instance, exhibits over 20,000 distinct crushed and distorted physical configurations in a waste stream.

```mermaid flowchart LR accTitle: Architecture of Greyparrot Waste Intelligence subgraph Input["Input Data"] A["Conveyor Streams\n(3 m/s, 40-50°C)"] B["100M+ Image Library\n(4 Continents)"] end subgraph Model["Waste Intelligence Core"] C["LSCVM Deep Learning Lens"] D["111 Material & Grade Categories"] end subgraph Output["Real-Time Facility Control"] E["Optical Sorters"] F["Robotic Arms"] G["Secondary Market Analytics"] end Input --> Model Model --> Output ```

Greyparrot engineered a dedicated Large-Scale Computer Vision Model (LSCVM) trained on an initial library of over 100 million images gathered across four continents. Sitting at the confluence of deep learning, computer vision, and material sciences, the model recognizes 99.9% of manufactured items categorized across 111 distinct material classes—spanning all varieties of plastics, cardboard, paper, ferrous, and non-ferrous metals, while distinguishing between food-grade and non-food-grade materials.

---

[10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) Inside the MRF: Creating the "Self-Driving" Recycling Facility

The ultimate fate of an item is decided not at household bins, but inside Material Recovery Facilities (MRFs), where municipal trucks deposit daily waste. High-capacity facilities process upward of 2,000 tons per day, while major global metropolitan areas like London or New York handle 7 million to 10 million tons of waste per day. MRFs rely on complex mechanical sorting sequences with up to seven distinct physical separation barriers to convert mixed waste into purified single-stream commodities.

Operating conditions on MRF sorting lines are severe: conveyor belts travel at 3 meters per second under poor ambient lighting, severe mechanical vibrations, heavy dust, and temperatures between 40°C and 50°C. Greyparrot mounts hardware directly above these belts, processing billions of object images daily at a fraction of the operating cost typical of major social networks—a necessity for an industry governed by fractions of a cent. Acting as an intelligent operating layer analogous to what Waymo does for conventional vehicles, the vision system feeds real-time telemetry into robotic arms and optical sorters to make plants fully automated and transparent.

---

[14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) DeepNest, Brand Accountability, and Packaging Physics

On June 24, Greyparrot released DeepNest, a public data platform designed to give consumer brands empirical visibility into the post-bin lifecycle of their packaging. Historically, industry recycling reports and UN assessments have frequently extrapolated wide conclusions from sample sets as small as 500 items—an approach Mitra compares to predicting the winner of a United States presidential election by surveying only five voters. DeepNest aggregates millions of empirical data points per brand to show whether expensive sustainability re-designs perform effectively in practice.

A primary example is the packaging transition executed by Sprite. Sprite historically packaged its beverage in green-tinted PET plastic before transitioning globally to clear, transparent PET. While both bottles are chemically identical PET, transparent bottles yield a 40% higher material recovery rate in recycling facilities. Seemingly cosmetic attributes such as bottle coloration, adhesive labels, full-wrap plastic sleeves, and cap compositions fundamentally dictate whether an item can be mechanically processed or ends up in a landfill.

- Greyparrot Waste Intelligence provides MRFs with higher secondary market pricing by guaranteeing verifiable purity thresholds on sorted commodity bales. - Packaging intelligence helps brand supply chains adapt to regulatory targets, NGO audits, and Gen Z consumer scrutiny. - Granular material analytics establish an empirical foundation for calculating carbon credits across discarded commodities. - Reliable waste-flow data gives policymakers the visibility required to enact viable environmental legislation rather than relying on flawed assumptions about recycling systems.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) Over 9 million people die annually from waste crisis-related fallout, surpassing the combined global death toll of AIDS, tuberculosis, and malaria. — Ambarish Mitra

> [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) Between $80 billion and $120 billion worth of economically viable materials are discarded globally each year due to inadequate processing infrastructure. — Ambarish Mitra

> [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) Decomposing or incinerated landfill waste generates nearly 10% of global greenhouse gas emissions—more than double the emissions generated by the entire commercial aviation sector. — Ambarish Mitra

> [04:07](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=247s) In the $23 trillion global consumption economy, consumer actions are tracked down to the millisecond by dozens of corporations targeting future purchasing intent. Yet the instant an item enters a bin, data collection vanishes. — Ambarish Mitra

> [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) Standard Large Multimodal Models and vision APIs fail completely when confronted with waste imagery because crushed, soiled, and torn objects do not exist in standard web training sets. — Ambarish Mitra

> [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) A standard Coca-Cola can, for instance, exhibits over 20,000 distinct crushed and distorted physical configurations in a waste stream. — Ambarish Mitra

> [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) Operating conditions on MRF sorting lines are severe: conveyor belts travel at 3 meters per second under poor ambient lighting, severe mechanical vibrations, heavy dust, and temperatures between 40°C and 50°C. — Ambarish Mitra

> [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) While both bottles are chemically identical PET, transparent bottles yield a 40% higher material recovery rate in recycling facilities. — Ambarish Mitra

> [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) Seemingly cosmetic attributes such as bottle coloration, adhesive labels, full-wrap plastic sleeves, and cap compositions fundamentally dictate whether an item can be mechanically processed or ends up in a landfill. — Ambarish Mitra

## Mentioned Ambarish Mitra, Greyparrot, Ai4, UN Plastics Treaty, United Nations, Coca-Cola, Amazon, DoorDash, Waymo, DeepNest, Sprite

## Video description Ambarish Mitra, Co-Founder, Greyparrot

Abstracts

We all consume. We all throw things away. But how much do we really know about what happens to our waste — or what it means for the planet?

Every year, over 2 billion tonnes of waste are processed through ageing, opaque systems with little accountability. The result? A broken link in the climate chain — and a massive, overlooked opportunity. The fight against waste is the fight against climate change.

In this keynote, Greyparrot co-founder Rish Mitra reveals how AI is lifting the lid on the world’s waste crisis. By digitising waste flows and unlocking real-time data from inside recycling facilities, AI is driving radical transparency and accelerating a circular economy. Just last year, Greyparrot’s AI analysed over 40 billion waste objects across more than 20 countries — giving us, for the first time, a clear picture of what we’re discarding, and how to fix it.

If you’ve ever wondered what really happens to your recycling — or what it will take to fix a broken system — this is the session you can’t afford to miss.

[Source](https://www.youtube.com/watch?v=iAuG8_aAu5o)

中文简报

Greyparrot追踪废弃物流，助力回收1200亿美元包装材料利用传送带计算机视觉填补废弃后的数据空白，对于实现分拣设施自动化，以及回收产生全球10%温室气体排放、价值1200亿美元的废弃商品至关重要。 Ambarish Mitra剖析了城市垃圾带来的巨大环境与经济成本：废弃物产生了全球近10%的温室气体排放，同时造成了价值数十亿美元的宝贵资源流失。他详细介绍了Greyparrot如何构建基于超过1亿张图像训练的专用大规模计算机视觉模型，以在高速传送带上识别涵盖111个材质类别的受压及受污物体。该技术作为材料回收设施的智能操作系统，并通过DeepNest平台提供包装可见性，不仅实现了分拣设施的自动化，还为消费品牌和政策制定者建立了基于经验数据的问责机制。消费后市政废弃物每年造成超过900万人死亡，并产生近10%的全球温室气体排放（超过商业航空的两倍），同时每年因分拣基础设施不足损失800亿至1200亿美元的可用材料。 (0:04) 全球23万亿美元的消费经济虽能将购买行为追踪至毫秒级，但废弃物一旦入桶便进入数据完全缺失的“阴影经济”，从有价资产沦为处理成本高昂的负债。 (4:07) 针对废弃物被压瘪、脏污导致通用视觉模型失效的痛点，Greyparrot基于超1亿张图像开发了大型计算机视觉模型（LSCVM），能识别111种材料与等级类别中99.9%的制成品。 (8:09) Greyparrot在传送带速度达3米/秒的严苛分选设施中部署AI以驱动机械臂与光学分选，并通过DeepNest数据平台提供海量实证数据，打破以往抽样过小的局限并推动品牌包装真实回收问责。 (10:49)

Reading summary

Greyparrot co-founder Ambarish Mitra argues that post-consumer waste forms an unmonitored 'shadow economy' that incurs massive environmental and economic costs, which can be mitigated by deploying specialized computer vision to automate sorting and enforce empirical producer accountability. The Economic and Climate Cost of Unmeasured Waste (0:04) Mitra highlights that everyday municipal waste represents an overlooked crisis compared to conventional climate imagery like oil rigs and industrial smokestacks. Fallout related to the global waste crisis claims more than 9 million lives each year—surpassing the combined tolls of AIDS, tuberculosis, and malaria. Furthermore, decomposing or incinerated landfill waste produces roughly 10% of global greenhouse gas emissions, more than double the emissions generated by the commercial aviation sector. This environmental toll is compounded by severe economic inefficiencies. Between $80 billion and $120 billion worth of viable materials are discarded globally each year due to insufficient recovery infrastructure, with United States delivery packaging alone shedding approximately $9.5 billion annually. Mitra contrasts the $23 trillion consumption economy—where consumer purchases are tracked down to the millisecond—with the post-disposal 'shadow economy,' where data tracking completely stops and physical objects abruptly flip from monetized assets into costly municipal liabilities. Computer Vision Adapted to Destroyed and Soiled Objects (8:09) Analyzing municipal waste presents unique computational challenges because no standard industry directly interfaces with billions of people generating trash weekly, and standard computer vision APIs fail on discarded items. Off-the-shelf vision models are trained on pristine web imagery and cannot identify crushed, soiled, or shredded materials; for instance, a single aluminum soda can exhibits over 20,000 distinct distorted shapes inside a waste stream. To solve this, Greyparrot engineered a dedicated Large-Scale Computer Vision Model (LSCVM) trained on over 100 million images sourced across four continents. Combining deep learning with material science, the system identifies 99.9% of manufactured items across 111 material and grade categories, including distinct plastic polymers among the 67 to 68 common variants, metals, cardboard, and critical separations between food-grade and non-food-grade items. Real-Time Sorting Automation Under Extreme Industrial Conditions (10:49) The actual recovery of municipal waste is determined at Material Recovery Facilities (MRFs), which process up to 2,000 tons per day per site, while metropolitan hubs like New York and London manage 7 million to 10 million daily tons. Sorting lines operate under severe conditions: conveyor belts move at 3 meters per second in dusty environments characterized by mechanical vibration, poor lighting, and temperatures between 40°C and 50°C. Greyparrot installs monitoring hardware directly above sorting belts to process billions of images at low operating cost. Operating as an intelligent operating layer—analogous to Waymo's software in autonomous vehicles—the system delivers real-time telemetry to robotic arms and optical sorters. This continuous monitoring enables facilities to automate complex sorting recipes, verify commodity bale purity for secondary markets, and support localized micro-recovery infrastructure such as containerized sorting pods. Empirical Packaging Intelligence and Physical Recyclability Limits (14:51) On June 24, Greyparrot launched DeepNest, a public data platform intended to replace historically tiny industry sample sizes—often as low as 500 items—with millions of real-world packaging data points. Mitra notes that brand claims and packaging re-designs frequently fail in practice because recyclability is governed by packaging physics and sorting mechanics rather than mere chemical identity. For example, Sprite's global transition from green-tinted PET to transparent clear PET yielded a 40% increase in facility recovery rates despite both bottles sharing identical chemical compositions. Cosmetic features such as full-wrap plastic sleeves, adhesives, cap materials, and dyes determine whether an item can be mechanically separated or diverted to a landfill, making continuous empirical waste data essential for packaging designers, brand compliance, and municipal carbon accounting. Mitra concludes that eliminating the data blind spot in the post-consumer waste stream is critical for both commercial recovery and public policy. Without empirical, facility-level data, initiatives like the UN Plastics Treaty and corporate sustainability commitments risk relying on flawed assumptions rather than the physical realities of global recycling infrastructure.

精炼阅读摘要

Greyparrot 联合创始人 Ambarish Mitra 指出，全球消费经济与废弃物处理之间存在巨大的数据断层，导致高额经济损失与碳排放；通过专为形变废弃物设计的计算机视觉模型，可实现垃圾分拣设施的自动化，并为品牌包装设计与环保政策提供实证数据支持。被忽视的消费后垃圾危机与数据断层 (0:04) Mitra 表示，主流气候叙事往往聚焦于石油钻井平台与工业烟囱，却忽视了日常消费品包装所引发的废弃物危机。全球每年因废弃物危机相关问题导致的死亡人数超过 900 万人，超过了艾滋病、结核病和疟疾的致死总和。此外，垃圾填埋分解或焚烧所产生的温室气体排放占全球近 10%，是整个商业航空领域排放量的两倍以上。由于缺乏完善的处理基础设施，全球每年丢弃的具有经济价值的材料达 800 亿至 120 亿美元，其中仅美国就因零售和外卖配送等环节每年损失约 95 亿美元的有价包装材料。这一问题的核心瓶颈在于价值 23 万亿美元的消费经济与废弃物“影子经济”之间的严重不对称。在消费经济中，企业对消费行为的追踪精细至毫秒级；然而，一旦商品被丢弃进垃圾桶，所有数据追踪便彻底中断。物品从拥有明码标价的资产，迅速转变为市政部门必须付费处置的零价值或负价值负债。Mitra 强调，泛指的“塑料”实际上涵盖 67 至 68 种常见变体（如 PET、HDPE、聚丙烯和聚苯乙烯），历史上的包装材料选择多由货架营销而非可回收性驱动。专用计算机视觉破解形变物体识别难题 (8:09) 废弃物处理行业需要直接对接全球数十亿人口每周产生的垃圾，传统人工分拣和通用机械系统无法实时应对这一庞大规模。通用大型多模态模型与标准视觉接口在废弃物图像面前往往完全失效，原因在于标准网络训练集并不包含被挤压、脏污、撕裂的物体形态。例如，一个普通的 Coca-Cola 易拉罐在废弃物流中就可能呈现出超过 20,000 种不同的受损形态。针对这一物理特性，Greyparrot 构建了大规模计算机视觉模型（LSCVM），其训练基础涵盖来自四大洲的超过 1 亿张图像。该模型结合了深度学习、计算机视觉与材料科学，能够识别 99.9% 的制成品，覆盖 111 个不同材料类别，包含各类塑料、纸板、纸张、黑色金属与有色金属，并能区分食品级与非食品级材料。材料回收设施的严苛工况与自动化改造 (10:49) 物品能否真正被回收，最终取决于材料回收设施（MRF）的处理能力。高产能的单体设施日均处理量可达 2,000 吨以上，而伦敦、纽约等主要都会区每天需处理 700 万至 1,000 万吨废弃物。回收设施通常依靠多达七道不同的物理分选屏障，将混合垃圾转化为高纯度的单一流向大宗原料商品。然而，MRF 分选流水线的实际作业环境极为恶劣：传送带以每秒 3 米的高速运转，伴随微弱照明、剧烈机械震动、高浓度粉尘，以及 40°C 至 50°C 的高温。Greyparrot 将硬件直接安装在传送带上方，类似 Waymo 对传统车辆的软硬件赋能，以远低于主流社交网络计算成本的支出，日均处理数十亿张物体图像，将实时遥测数据反馈给光学分选机和机械臂，推动回收工厂向透明化和全面自动化转型，同时确保分选包达到可验证的纯度门槛，从而提升二级市场售价。微观包装物理学与宏观政策数据支持 (14:51) 行业传统的回收报告和联合国评估有时仅基于 500 个样本的小数据集推导结论，Mitra 将其比作仅调查五名选民就预测美国总统大选结果。Greyparrot 随后推出了 DeepNest 公共数据平台，通过汇总单个品牌数以百万计的经验数据，检验品牌高成本的可持续包装重塑在实际回收流中是否奏效。包装的细微物理属性直接决定其能否被机械回收。例如，Sprite 将其饮料瓶从传统的绿色 PET 改为透明 PET，尽管两种材质在化学成分上完全相同，但透明瓶在回收设施中的材料回收率直接提升了 40%。瓶身颜色、粘合剂标签、全包裹塑料套标和瓶盖成分，均对物品最终进入循环流还是填埋场起着决定性作用。这些实证流向数据为品牌应对监管目标、市政区域制定切实可行的环保法规，以及精确核算废弃商品碳信用提供了基础支持。 Greyparrot 试图通过部署在工业传送带上的专用计算机视觉系统，填补消费链条末端的数据盲区，为废弃物分拣自动化、包装可回收性验证与政策制定提供实证支撑；然而，技术落地仍高度依赖高通量回收设施的严苛硬件适应性，以及品牌与市政体系在包装标准和基础设施上的协同改变。

Detailed notes — zh

## Background 本内容由 Ai4 于 2025 年 8 月 28 日在 YouTube 平台发布，记录了 Greyparrot 联合创始人 Ambarish Mitra 的主题演讲。视频简介指出，全球每年有超过 20 亿吨垃圾流经不透明且老化的处理系统，演讲揭示了 Greyparrot 如何通过 AI 计算机视觉技术将垃圾流数字化，在回收设施中解锁实时数据，从而推动废弃物透明度并加速循环经济发展。 ## TL;DR Ambarish Mitra 探讨了长期被气候讨论所忽视的全球垃圾危机，指出被丢弃的消费包装跌入了缺乏数据追踪的“影子经济”，造成了惨重的健康与经济损失。为此，Greyparrot 开发了针对变形破损物体的超大规模计算机视觉模型（LSCVM），部署在恶劣的材料回收设施（MRF）流水线上，将传统机械回收厂转变为类似 Waymo 赋能的“无人自驾驶工厂”。通过毫秒级实时分类与品牌归因平台 Deepnest，该技术不仅帮助回收厂提纯物料以提高利润，还撕下了品牌的漂绿伪装，为包装设计改良、政策制定及碳信用核算提供了前所未有的实证数据底座。 ## Atmosphere & style 演讲展现出一种兼具行业紧迫感与硬核科技理性的极客气质。Ambarish Mitra 的语调沉稳笃定且带着犀利的幽默，尤其在对比消费经济的精密追踪与垃圾处理的蛮荒现状，或调侃大语言模型在脏乱废弃物面前束手无策时，观点直击行业痛点。现场营造出一种打破认知盲区的震撼感，带领听众从光鲜的消费社会一举下潜至充斥着粉尘与剧烈震动的垃圾传送带，完成了从环保宏大叙事到严谨工业落地的精彩转折。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) 垃圾危机是气候变化中被严重低估的战场：填埋与焚烧产生的温室气体占全球总排放量的近 10%（超过航空业的两倍），每年还造成 800 亿至 1200 亿美元高价值原材料的直接损失。 - [04:07](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=247s) 塑料包含近 70 种复杂细分聚合物变体，品牌为了追求视觉营销效果导致包装材料过度复杂化，推进材料的简化与标准化是提升全球回收率的关键前提。 - [05:54](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=354s) 消费端拥有高达 23 万亿美元且被精密追踪的数字经济，但物品被扔进垃圾桶后便瞬间跌入缺乏数据追踪、被视为零价值甚至负资产的“影子经济”断层。 - [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) 废弃物流水线上物料高度破损与变形（单个易拉罐可呈现超 2 万种形变），且缺乏公共互联网数据，使得通用大模型（如 OpenAI、Gemini）在此类极端物理场景下完全失效。 - [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) Greyparrot 打造的大规模计算机视觉模型（LSCVM）结合深度学习与材料科学，涵盖 111 个材料大类，可对 99.9% 严重变形挤压的人造物品进行精准识别与分类。 - [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) 材料回收设施（MRF）是决定垃圾走向循环还是填埋的枢纽，引入嵌入式 AI 视觉系统后，传统机械工厂得以像 Waymo 赋能汽车一样蜕变为“无人自驾驶工厂”。 - [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) 废弃物检测系统在粉尘、高温和 3 米/秒的高速传送带等极端工况下，以仅为主流社交网络千分之一的超低推理成本，实现了每日数十亿级的工业图像实时处理。 - [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) 基于海量入桶后真实数据的 Deepnest 平台撕掉了品牌的“漂绿”伪装，例如实证表明将绿色雪碧瓶改为透明瓶身可使回收转化率直接提升 40%。 - [17:37](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=1057s) 废弃物智能全面赋能全生态链：帮助回收设施提纯单流物料以获取溢价，为品牌提供真实的包装减塑洞察，并为碳信用核算与政府环境监管立法提供高颗粒度的工业实证数据。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) 垃圾危机：被气候讨论忽视的隐形战场 Greyparrot 联合创始人 Ambarish Mitra 首先回顾了自己的创业背景：在 2012 年，他创立了世界上首个视觉搜索引擎，致力于将物理世界中所有形态完好的物体绘制成数字模型。然而到了 2019 年，他受到启发，意识到世界上还有一个完全被所有人忽略的庞大领域亟待探索——那就是废弃物领域。每当人们谈论气候变化时，脑海中浮现的往往是钻油平台、高耸冒烟的烟囱以及浓重的工业污染。极少有人会联想到人类每天制造的城市生活垃圾，或是我们日常丢进垃圾桶里的物品。然而，除了马斯洛需求层次理论中所描述的食物、住所、衣物和爱等基本需求之外，人类还有一个普遍共同点：每个人每天都在消费，而这些消费品几乎都带有包装。无论一个人生活在世界多么偏远的角落，全球化已经让所有物品都包裹在某种形式的包装之中。大多数人以为把垃圾扔进垃圾桶就尽到了责任，整个旅程便就此终结；但事实上，这些物品和材料才刚刚开启一段漫长且鲜为人知的旅程，而这段旅程伴随着严重的破坏性后果。 - **健康代价**：由于垃圾危机引发的污染渗透进水源与社会（尤其在发展中国家），每天导致超过 900 万人丧生，这一数字超过了艾滋病、肺结核和疟疾致死人数的总和。 - **经济损失**：每年有价值 800 亿至 1200 亿美元具有即时经济价值的原材料被直接丢弃，原因仅仅是缺乏足够的处理和评估基础设施；仅在美国，从零售商、DoorDash 或亚马逊等平台购买商品所丢弃的材料与包装价值就高达 95 亿美元。 - **气候排放**：填埋场中腐烂、焚烧的垃圾所产生的温室气体排放占全球总排放量的近 10%，这一排放规模是常被公众口诛笔伐的航空业总排放量的两倍以上。 --- ### [04:07](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=247s) 联合国塑料条约谈判与塑料复杂性的本质 Mitra 提到他前一天刚从日内瓦飞抵现场。彼时正值联合国塑料条约第六轮谈判（INC 谈判），虽然主流媒体关注有限，但这关乎全球塑料治理的关键转折。此前五轮谈判均未取得实质性成果，各方正翘首以盼能够在两天后的 8 月 14 日达成具有约束力的协议，强化“生产者责任延伸（EPR）”机制，让品牌为其制造的包装承担更多法律与经济责任。日内瓦的核心议题不仅在于提高回收率或向生产商征税，更聚焦于源头减量——探讨人类是否真的需要如此大量的塑料。此外，大众日常将塑料统称为单一类别，这极大掩盖了材料科学的复杂性。塑料是由复杂的聚碳酸酯及其他高分子聚合物组成的系统，日常生活中存在诸如 PET、聚乙烯（PE）、聚苯乙烯（PS）、高密度聚乙烯（HDPE）等 67 至 68 种细分变体。这些材料的回收难度各不相同，而之所以催生出如此繁复的种类，完全是品牌为了吸引眼球、满足营销视觉需求和颜色设计所致。若能对材料进行简化与标准化，全球回收率将迎来爆发式改善。 --- ### [05:54](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=354s) 从 23 万亿消费经济跌落至无人问津的“影子经济” 为了向在场听众阐述垃圾数据化的迫切性，Mitra 展示了一个商业漏斗图：在粉色圆圈标示的消费端，全球每一毫秒都在追踪、分析并交易消费者的数字足迹。无论在欧美、印度还是中东，至少有 10 家公司时刻掌握消费者的行踪与下一步购买意向。这是一个规模高达 23 万亿美元的全球消费经济，因为背后蕴藏着巨大的商业回报，所以技术被推向了极其精密的高度。然而，一旦消费者将物品丢入垃圾桶，这个物品便瞬间跌入所谓的“影子经济（Shadow Economy）”。社会大众不再将其视作具有价值的材料资产，而是彻底打入冷宫。以塑料瓶为例：企业花钱购买它作为容器，它本应具有公允价格；但落地垃圾桶后，它不仅被视作零价值资产，甚至成了负资产——政府必须耗费财政公款去填埋或销毁它。人们普遍缺乏将废弃物视为“可在循环经济中再利用的同一物体”的思维模式。这种断层的根源，在于长期以来没有任何商业力量尝试去追踪和交易废弃物数据。Greyparrot 的诞生正是为了在庞大尺度上绘制消费后世界的数字地图，精确追踪从巧克力包装纸到 100 英寸电视机包装箱的每一个丢弃物，甚至能做到实时量化评估“昨天整个纽约市的废弃物处理表现”。 ```mermaid flowchart TD accTitle: 消费经济与影子经济的流转断层 accDescription: 展示物品从高精追踪的消费端跌入无人问津的废弃物处理端的断崖式转变 subgraph Consumption["23 万亿美元消费经济"] A["高精数据追踪 (每毫秒行为预测/精准营销)"] --> B["商品购买与使用 (具有明确商业与材料价值)"] end subgraph Shadow["废弃物影子经济 (断层阶段)"] B -. "丢入垃圾桶 (被视为零价值或负资产)".-> C["传统处置流向 (填埋/焚烧/环境污染)"] B ==> D["Greyparrot AI 视觉感知层 (材质分类/纯度检测/品牌归因)"] end subgraph Circular["闭环循环经济"] D --> E["MRF 智能分选与纯单流分离"] E --> F["高纯度二级原料市场交易"] F --> B end ``` --- ### [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) 废弃物智能的突破：LSCVM 大规模计算机视觉模型垃圾处理之所以长期被视为无法解决的难题，原因在于没有任何一个行业的物理体量能与废弃物相提并论——全球 70 亿人口每周都会制造一袋半到两袋的垃圾，没有任何现有机械或人工能在真实时间内完全处理这种吞吐量。面对这一行业痛点，唯有借助人工智能才能在经济上可行地理解和重塑这一链路。为此，Greyparrot 构建了一套大规模计算机视觉模型（LSCVM）。这套模型能够识别 99.9% 的人造物品，其核心技术壁垒在于：即便物品被严重挤压、破损变形，它依然能够精准辨认。Mitra 幽默地指出，在理想的零售环境、家庭或社交网络中，一罐可口可乐拥有标准完好的外观；但到了废弃物流水线上，由于遭受暴力挤压和破损，一个易拉罐能呈现出 20,000 多种截然不同的变形形态。对于在场迷恋 LLM（大语言模型）的技术听众，他直言：“如果你从我们的数据集中随便取样扔给 Gemini 或 OpenAI，得到的结果会是零——因为这些高度破碎、污染严重的数据根本不存在于已知的公共互联网上。” - **多学科融合**：废弃物智能（Waste Intelligence）站在计算机视觉、大规模深度学习与材料科学的交叉点上。 - **庞大数据积累**：前期从四大洲采集了约 1 亿张初始图像，建立起跨品类的单一通用识别镜组。 - **细分材料颗粒度**：系统目前涵盖 111 个材料大类，贯穿各类塑料、纸张、硬纸板、黑色金属及有色金属等。 --- ### [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) 材料回收设施（MRF）与无人自驾驶工厂决定一件消费品宿命的真正场所不是家里的分类垃圾桶，而是被称为材料回收设施（Material Recovery Facility，简称 MRF）的工业枢纽。城市垃圾车将废弃物从居民区收集后统一运往这里。Mitra 比喻道，一个包装最终是升入“天堂”（再利用、高质回收循环）还是堕入“地狱”（填埋、焚烧），完全由 MRF 决定。在美国，大型 MRF 每天可处理约 2,000 吨垃圾，而纽约或伦敦这样的特大城市，全城系统每日吞吐的垃圾更是高达 700 万至 1,000 万吨。在传统的 MRF 内部，物料要在混合状态下通过多达七道机械隔离线逐步筛选成单一组分。Greyparrot 的嵌入式摄像头系统对传送带上的每一件废弃物进行毫秒级实时分类，不仅能识别材质，还能区分食品级与非食品级塑料，并将纯度数据即时反馈给工厂操作人员。操作员根据设定的阈值配方动态控制产线，使原本粗放的物理工厂变成半自动化乃至全自动化的运转系统。这种检测信号还可以直接联动下游的分选机械臂和光学分选机。Mitra 将这一变革类比为自动驾驶：“这就如同 Waymo 对传统汽车的赋能——原本冰冷机械的汽车变成了自动驾驶汽车；引入 Greyparrot 视觉感知系统后，传统机械回收厂蜕变成了‘自动驾驶工厂’，实现 100% 的透视度，没有任何一个经过的物体会被遗漏。” - **极端的工业视觉工况**：传送带移动速度高达 3 米/秒，现场伴随持续剧烈震动、光照极度昏暗、充斥粉尘污垢，车间环境温度常常达到 40 至 50 摄氏度，与干净明亮的零售环境完全相反。 - **吞吐量与成本优势**：Greyparrot 每日处理的图像体量堪比全球主流社交网络（单日处理数十亿级图像），但推理运行成本仅为其千分之一，因为废弃物回收是一个“对每吨几美分都精打细算”的薄利工业。 --- ### [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) 全球化实践落地：从韩国智能集装箱到 Deepnest 品牌数据库 Greyparrot 现已成为全球领先回收设施制造商的核心合作伙伴，并向世界各地提供开放 API。在被公认为回收与循环基础设施最先进的韩国，该技术不仅部署于集中式工厂，还被直接装配到城市各个街区的集装箱式分布式回收机中，通过内置机械臂在源头取代传统街头分类桶，实现进箱物料的即时精准分流。与此同时，在处理了数十亿废弃物后，团队迎来了一项关键突破：他们不仅能识别物理材质，还能反向识别其包装背后的品牌归属。2024 年 6 月 24 日，Greyparrot 正式上线了面向公众的废弃物大数据平台——Deepnest。这一系统撕掉了许多大品牌“漂绿（Greenwashing）”的伪装。在此之前，即便是联合国编写的气候或塑料调查报告，其样本量往往也仅有区区 500 个，Mitra 讽刺这种做法“无异于只调查了五个人就宣布美国总统大选结果”；而 Deepnest 汇集了数以亿计的真实 post-bin（入桶后）全量数据，让欧美各大消费巨头清醒地看到自己每年砸下数百万美元设计的包装在现实中究竟流向了何方。 | 包装形态与材质 | 细分材质类型 | 回收率与商业影响 | 行业启示 | |:--- |:--- |:--- |:--- | | **传统绿色雪碧瓶** | PET（聚对苯二甲酸乙二醇酯） | 较低（因颜色色素限制降级利用） | 染色导致再生料价值缩水，回收端阻力大 | | **透明白色雪碧瓶** | PET（同等化学材质） | **回收转化率直接提升 40%** | 单纯去色化即可极大提升循环经济收益 | 通过雪碧瓶的案例，Mitra 强调包装的可回收性绝不仅仅取决于材料科学本身，瓶身颜色、热收缩套标材质、纸质标签和瓶盖分离度都起着生死攸关的作用，而这些实证数据在过去完全是一片空白。 --- ### [17:37](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=1057s) 废弃物智能如何赋能全生态链在演讲结语中，Mitra 总结了废弃物智能为产业链各方带来的实质赋能： - **对材料回收设施（MRF）**：直接带来经济增益。回收是一门商业生意，AI 质控帮助他们分离出高纯度单流塑料，使其能够在二级塑料市场以更高的溢价脱手，大幅提升了回收厂的盈利能力。 - **对消费品牌（Brand Owners）**：提供基于真实数据的包装供应链洞察。面对来自 NGO 和 Z 世代消费者的减塑问责压力，品牌能依据 post-bin 数据作出正确的包装轻量化与材料替代决策。 - **对碳信用市场**：由于几乎所有大宗消费品最终都会转化为废弃物形态，系统积累的全生命周期物料追踪数据可为核算真实碳抵消与碳信用提供权威客观的依据。 - **对监管机构与政策制定者**：呼吁立法机构在制定或修改环保法令、征税细则前，必须依赖高颗粒度的工业实证数据，切忌在缺乏对回收工业现实认知的情况下拍板决策。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) 由于垃圾危机引发的污染渗透进水源与社会（尤其在发展中国家），每天导致超过 900 万人丧生，这一数字超过了艾滋病、肺结核和疟疾致死人数的总和。 — Ambarish Mitra > [00:04](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=4s) 填埋场中腐烂、焚烧的垃圾所产生的温室气体排放占全球总排放量的近 10%，这一排放规模是常被公众口诛笔伐的航空业总排放量的两倍以上。 — Ambarish Mitra > [05:54](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=354s) 企业花钱购买它作为容器，它本应具有公允价格；但落地垃圾桶后，它不仅被视作零价值资产，甚至成了负资产——政府必须耗费财政公款去填埋或销毁它。 — Ambarish Mitra > [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) 在理想的零售环境、家庭或社交网络中，一罐可口可乐拥有标准完好的外观；但到了废弃物流水线上，由于遭受暴力挤压和破损，一个易拉罐能呈现出 20,000 多种截然不同的变形形态。 — Ambarish Mitra > [08:09](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=489s) 如果你从我们的数据集中随便取样扔给 Gemini 或 OpenAI，得到的结果会是零——因为这些高度破碎、污染严重的数据根本不存在于已知的公共互联网上。 — Ambarish Mitra > [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) 一个包装最终是升入“天堂”（再利用、高质回收循环）还是堕入“地狱”（填埋、焚烧），完全由 MRF 决定。 — Ambarish Mitra > [10:49](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=649s) 这就如同 Waymo 对传统汽车的赋能——原本冰冷机械的汽车变成了自动驾驶汽车；引入 Greyparrot 视觉感知系统后，传统机械回收厂蜕变成了‘自动驾驶工厂’，实现 100% 的透视度，没有任何一个经过的物体会被遗漏。 — Ambarish Mitra > [14:51](https://www.youtube.com/watch?v=iAuG8_aAu5o&t=891s) 即便是联合国编写的气候或塑料调查报告，其样本量往往也仅有区区 500 个，无异于只调查了五个人就宣布美国总统大选结果。 — Ambarish Mitra ## Mentioned Ambarish Mitra, Greyparrot, 联合国塑料条约 (INC 谈判), 生产者责任延伸 (EPR), DoorDash, 亚马逊, Gemini, OpenAI, Waymo, Deepnest, 材料回收设施 (MRF), 大规模计算机视觉模型 (LSCVM), 可口可乐, 雪碧 ## Video description Ambarish Mitra, Co-Founder, Greyparrot Abstracts We all consume. We all throw things away. But how much do we really know about what happens to our waste — or what it means for the planet? Every year, over 2 billion tonnes of waste are processed through ageing, opaque systems with little accountability. The result? A broken link in the climate chain — and a massive, overlooked opportunity. The fight against waste is the fight against climate change. In this keynote, Greyparrot co-founder Rish Mitra reveals how AI is lifting the lid on the world’s waste crisis. By digitising waste flows and unlocking real-time data from inside recycling facilities, AI is driving radical transparency and accelerating a circular economy. Just last year, Greyparrot’s AI analysed over 40 billion waste objects across more than 20 countries — giving us, for the first time, a clear picture of what we’re discarding, and how to fix it. If you’ve ever wondered what really happens to your recycling — or what it will take to fix a broken system — this is the session you can’t afford to miss. [Source](https://www.youtube.com/watch?v=iAuG8_aAu5o)

Source quotes

[0:04] Ambarish Mitra: 由于垃圾危机引发的污染渗透进水源与社会（尤其在发展中国家），每天导致超过 900 万人丧生，这一数字超过了艾滋病、肺结核和疟疾致死人数的总和。 [0:04] Ambarish Mitra: 填埋场中腐烂、焚烧的垃圾所产生的温室气体排放占全球总排放量的近 10%，这一排放规模是常被公众口诛笔伐的航空业总排放量的两倍以上。 [5:54] Ambarish Mitra: 企业花钱购买它作为容器，它本应具有公允价格；但落地垃圾桶后，它不仅被视作零价值资产，甚至成了负资产——政府必须耗费财政公款去填埋或销毁它。 [8:09] Ambarish Mitra: 在理想的零售环境、家庭或社交网络中，一罐可口可乐拥有标准完好的外观；但到了废弃物流水线上，由于遭受暴力挤压和破损，一个易拉罐能呈现出 20,000 多种截然不同的变形形态。 [8:09] Ambarish Mitra: 如果你从我们的数据集中随便取样扔给 Gemini 或 OpenAI，得到的结果会是零——因为这些高度破碎、污染严重的数据根本不存在于已知的公共互联网上。 [10:49] Ambarish Mitra: 一个包装最终是升入“天堂”（再利用、高质回收循环）还是堕入“地狱”（填埋、焚烧），完全由 MRF 决定。 [10:49] Ambarish Mitra: 这就如同 Waymo 对传统汽车的赋能——原本冰冷机械的汽车变成了自动驾驶汽车；引入 Greyparrot 视觉感知系统后，传统机械回收厂蜕变成了‘自动驾驶工厂’，实现 100% 的透视度，没有任何一个经过的物体会被遗漏。 [14:51] Ambarish Mitra: 即便是联合国编写的气候或塑料调查报告，其样本量往往也仅有区区 500 个，无异于只调查了五个人就宣布美国总统大选结果。

Source quotes

[0:04] Ambarish Mitra: Over 9 million people die annually from waste crisis-related fallout, surpassing the combined global death toll of AIDS, tuberculosis, and malaria. [0:04] Ambarish Mitra: Between $80 billion and $120 billion worth of economically viable materials are discarded globally each year due to inadequate processing infrastructure. [0:04] Ambarish Mitra: Decomposing or incinerated landfill waste generates nearly 10% of global greenhouse gas emissions—more than double the emissions generated by the entire commercial aviation sector. [4:07] Ambarish Mitra: In the $23 trillion global consumption economy, consumer actions are tracked down to the millisecond by dozens of corporations targeting future purchasing intent. Yet the instant an item enters a bin, data collection vanishes. [8:09] Ambarish Mitra: Standard Large Multimodal Models and vision APIs fail completely when confronted with waste imagery because crushed, soiled, and torn objects do not exist in standard web training sets. [8:09] Ambarish Mitra: A standard Coca-Cola can, for instance, exhibits over 20,000 distinct crushed and distorted physical configurations in a waste stream. [10:49] Ambarish Mitra: Operating conditions on MRF sorting lines are severe: conveyor belts travel at 3 meters per second under poor ambient lighting, severe mechanical vibrations, heavy dust, and temperatures between 40°C and 50°C. [14:51] Ambarish Mitra: While both bottles are chemically identical PET, transparent bottles yield a 40% higher material recovery rate in recycling facilities. [14:51] Ambarish Mitra: Seemingly cosmetic attributes such as bottle coloration, adhesive labels, full-wrap plastic sleeves, and cap compositions fundamentally dictate whether an item can be mechanically processed or ends up in a landfill.

Original transcript

[0:04] Hi everyone. Uh my name is Amirish Mitra [0:08] uh co-founder of Gray Parrot. Uh before [0:13] uh founding Gray Parrot, I created the [0:15] world's uh first visual search engine in [0:18] 2012 [0:19] where we were mapping [0:23] everything we see in the physical world [0:26] in a perfect shape. And then in 2019 [0:31] I got an inspiration that actually [0:34] there's whole another world which no [0:36] one's talking about which I'll talk [0:39] about a little more in details. [0:42] Whenever we think of climate change the [0:46] image that appears in our head is oil [0:50] rigged and stacked up chimneys [0:53] um a lot of pollution. [0:56] Very rarely whenever I speak to people [0:59] the actual municipal waste created by [1:02] humans where things we throw in our bins [1:05] regularly [1:07] comes to our minds but we don't realize [1:11] that actually [1:13] uh uh as humans sorry I'm clicking [1:18] ah perfect thank you we don't realize [1:21] that [1:23] uh all of us [1:25] leave a very large footprint in society [1:29] every day when we consume lot of things. [1:33] something common about humans beside [1:36] what we read in Maslo's hierarchy of [1:38] needs and love and other things food [1:40] shelter clothing each one of us consumes [1:44] something or other every day which a [1:46] part of it gets disposed in some form of [1:49] packaging you know no matter how remote [1:53] uh you live in the world globalization [1:55] has led to everything coming in some [1:58] form of packaging today and and even [2:02] though You feel you've done your bit by [2:07] throwing it in the in the bins and you [2:09] feel that the journey ends there. [2:12] There's a very long journey to these [2:15] products and materials which very few [2:18] people know about and there's very [2:19] little actual transparency what happens. [2:24] Um and one of the things is in that [2:26] journey that journey is not without its [2:30] own harmful outcomes. [2:34] So you may not realize that just as a [2:38] result of waste crisis [2:40] uh more than 9 million people lose their [2:43] lives every day. That's more than uh [2:46] AIDS, tuberculosis and malaria uh put [2:49] together. how wasted infiltrating our [2:53] waters and societies and many parts of [2:55] developing worlds. [2:58] More than 80 to$120 billion worth in [3:01] financial value materials are thrown [3:04] which have immediate measurable economic [3:08] value uh because there's not enough [3:10] infrastructure to process and value it. [3:14] In US alone, that number is around $9.5 [3:17] billion uh of worth of materials and [3:20] packaging every time you buy things from [3:22] a retail or Door Dash or Amazon and [3:25] other things. And lastly, [3:29] just a lot of our waste that ends up in [3:33] landfill either gets, you know, either [3:35] rots there or there's incineration or it [3:38] gets burned. And that leads to almost [3:42] 10,000 10% of greenhouse emissions which [3:45] is more two times more than uh entire [3:48] aviation uh industry which so many [3:51] people talk about so frequently [3:54] uh as a as a big contributor. This is [3:56] just to give you a macro view uh the [4:01] importance of waste as a category in uh [4:04] global greenhouse emissions. [4:07] I flew in yesterday from uh Geneva. It's [4:11] a very important week. Uh it's the sixth [4:14] UN plastics treaty uh going on right [4:18] now. It may not be as much in mainstream [4:21] news in our circles here, but it's a [4:24] very important discussion. We've had [4:26] five roundts in last few years without [4:30] any outcome and we are really hoping for [4:33] a positive outcome which brings uh makes [4:37] producers more responsible for the [4:39] packaging it's creating [4:41] uh and a lot more accountability and [4:43] fingers crossed the result is on uh on [4:47] the 14th uh day after tomorrow. But one [4:50] of the big discussions going on there is [4:53] to not just improve recoverability [4:57] and bring some kind of a plastic stacks [5:00] uh to producers but also uh reduction [5:05] that do we need uh this much amount of [5:08] uh plastic and lastly as you know we [5:12] oversimplify the phrase by using [5:14] plastic. It's it's a chemic it's a very [5:18] complex chemical polycarbonate. So and [5:22] PETP [5:24] uh polyethine polystyrene uh HDP these [5:28] are all they're like 6768 variants of [5:31] products which you use uh every day in [5:34] many forms uh simplifying and [5:36] standardizing them they're not all of [5:38] them recover equally well and the reason [5:40] we have so many variants is due to [5:43] colors and marketing goals brands have [5:45] had uh and to get our attention. So [5:48] simplifying that would really really [5:51] help with recovery. [5:54] And this is my last point before I come [5:56] to the AI part. Why framing this is [5:59] important for you guys. Um what do you [6:02] see in this uh funnel here is uh where [6:06] the pink circle is that's where the [6:09] consumer is. all of us in this room uh [6:13] between in every economy whether it's [6:15] the US or the UK or India or Middle East [6:18] anywhere um the consumers are being [6:23] traced and tracked and traded every [6:25] millisecond at least 10 companies know [6:28] where you are what you're about to do [6:30] and your information is being traded and [6:32] your next intention what you'll be [6:33] purchased targeted it's so measured and [6:36] it's so advanced right and that's what [6:39] we call the consump consumption economy [6:41] because there's money in it, right? And [6:42] that's a $23 trillion global economy. [6:46] The moment you consume something and you [6:48] throw it in the bin, we call it the [6:50] shadow economy and no one cares about it [6:53] because you think of it as waste. You [6:57] don't think of it as material asset. [6:59] That plastic bottle, someone bought it [7:02] as a container for the goods you [7:03] consumed. It had a price. when you threw [7:06] it in the bin, now it is being not only [7:09] seen as a zero value asset, it's a [7:11] negative value because it'll cost the [7:13] government some money to dispose it. [7:15] Instead of the mindset that it's not a [7:17] waste, it's the same object and it could [7:19] come back into the circular economy. And [7:22] one of the reasons it's not happening [7:25] that way because of lack of companies [7:29] trying to trace and trade it. So we [7:32] started gray parrot to actually map the [7:34] postconumption world at an enormous [7:37] scale which we'll talk about where we [7:39] wanted to trace and track [7:42] everything from a chocolate wrapper to a [7:45] 100in television box which every human [7:48] throws in every city. When I say every, [7:52] I literally mean every. You know today, [7:56] just to give an example, how did New [7:58] York perform yesterday as in its waste? [8:01] We have that level of data. Uh, and [8:04] we'll explain in a moment why. [8:09] So, this is what I was framing up to. [8:11] Imagine a world where every piece of [8:14] waste is measured and valued. And the [8:17] reason this problem was unsolvable and I [8:20] want you to genuinely reflect on it. [8:23] Imagine there are so many industries [8:25] each one of you represent. But there is [8:28] no industry at the scale of waste [8:30] because it connects 7 7 billion people [8:34] every day on a weekly basis producing a [8:37] bag a bag and a half of waste. There is [8:40] no system, machinery, human labor [8:43] available to get rid of it in real time. [8:47] It's a large problem and it's very easy [8:49] to criticize the industry and it was [8:51] prime for a scale of artificial [8:54] intelligence to understand and solve it. [8:57] Uh there was no other way to do it [8:59] because it'll be so much more expensive. [9:02] So we created a large scale computer [9:06] vision model we call it LSCVM [9:09] which recognizes 99.9% [9:12] of things humans make but in a shape and [9:17] form even when it's destroyed. So in an [9:20] ideal world, a Coke can looks like a [9:23] Coke can whether you seeing it in a [9:26] retail environment or in a garden [9:28] environment or in your home or in any [9:31] entertainment form. But in the waste [9:33] world, a Coca-Cola can has more than [9:36] 20,000 variations of faces because it's [9:39] crushed. It's difficult. It's a very [9:42] large and complex problem to solve. And [9:45] those of you are big fans of LLN's in [9:48] this room, try and use any sample of our [9:52] data set and run it through Geminis and [9:54] Open AIS of the world. You'll get a zero [9:57] result because this data isn't available [10:00] on known internet. [10:04] Here comes [10:06] the concept of waste intelligence. It's [10:08] at the intersection of computer vision, [10:11] large scale deep learning, and material [10:14] sciences. It has taken uh around 100 [10:18] million initial images to create the [10:20] library from four different continents [10:24] uh to create one single lens which [10:27] recognizes every literally everything [10:30] 111 categories of material types from [10:34] all forms of plastic, paper, cardboard, [10:36] ferrris, non-ferris, metals everything [10:39] that leads to what you typically consume [10:43] either the main product or the packaging [10:45] itself and these are snapshots [10:49] uh that are going in conveyor belts. [10:51] These are places called material [10:53] recovery facilities in the typical waste [10:56] advanced infrastructure world that is [10:59] mostly US uh Europe parts of South Asia, [11:03] Middle East and parts of South America [11:05] where we have morphs which is called [11:08] material recovery facilities where most [11:10] of city's waste when it's collected from [11:12] your bin into trucks that end up here. [11:16] Why this place is important not your [11:18] bin? The destiny of a product is decided [11:21] here not at home. So whether your [11:24] product will be going to [11:27] heaven in terms of some kind of a [11:30] recovery or repurpose or hell which is [11:32] some kind of incineration or you know [11:35] get into landfill or have some kind of [11:37] an afterlife uh hopefully uh is decided [11:41] here. So, and some morphs produce like [11:45] the largest one in the US produces sorry [11:47] processes around 2,000 tons uh of waste [11:51] per day and cities like New York or [11:55] London they process anything between 10 [11:59] million to 7 million to 10 million tons [12:01] of waste uh per day. Uh so [12:06] once we recognize everything that's [12:09] happening so this is what a typical [12:10] morph looks like all all objects are [12:15] classified in real time you know by [12:18] knowing what it is also classified by [12:21] material whether it's food grade [12:23] non-food grade the plant owner gets the [12:27] classification of what's coming through [12:29] it's purity level they need to they need [12:31] to know a certain level of threshold [12:33] control a recipe. They can control the [12:36] plant, make it semi-automated and then [12:38] they make it into 100% single streams of [12:42] different plastics so that they can [12:43] eventually sell it in the secondary [12:46] market. There are various lines almost [12:50] every plant has seven different [12:51] barriers. Everything comes into a mixed [12:54] from your mixed garbage then becomes [12:56] into single stream. How are they [12:58] quality? how the quality control happens [13:01] before or after. Also, the signals from [13:05] the machine is sent to um robotic arms, [13:10] optical sortters. So, think of it as a [13:13] mechanical infrastructure which didn't [13:14] have a brain. So, this is almost like [13:17] what Whimo does to a car, right? [13:21] A normal mechanical car becomes a [13:22] self-driving car. So with gray parrot [13:24] camera systems, a plant becomes a [13:27] self-driving plant with 100% [13:30] transparency and not a single object [13:33] being undetected. [13:35] [Music] [13:36] Uh keeping in mind from a computer [13:38] vision challenge perspective, [13:41] these belts move at 3 m a second. It's [13:44] very dirty. There's poor lighting. [13:47] It's constantly vibrating. It's like 40 [13:50] 50° temperature. So it's not a it's like [13:53] absolutely opposite of a retail world. [13:55] So it's very challenging uh from a AI [13:59] perspective. So it's a lot of incredible [14:01] amount of work and those curious about [14:03] on the technical side feel free to talk [14:05] to me later. But it's a breakthrough. It [14:09] process we process probably same images [14:11] amount of images as most social networks [14:13] in the world. Uh like the number goes [14:16] into billions per day. and uh we do it [14:19] at 1,000th of a cost to most social [14:22] networks. So because the industry is not [14:25] advertising le uh it's it's second it's [14:28] a second decimal scent conscious [14:30] industry. [14:32] Moving forward, this is just a quote [14:35] from uh one of our clients is the [14:36] largest plant builder in the world has [14:38] best built most of the plants in America [14:41] who are actually talking about uh how [14:44] this is a gamecher for the industry and [14:47] bringing large scale transparency. [14:51] I'll show you another use case. This is [14:54] in uh South Korea where the same system [14:58] is now being South Korea is the world's [15:00] most advanced place for recovery and [15:03] recycling. So where robots are being [15:06] used in in containers literally like [15:10] containers ship style containers in [15:12] cities in various places where our tech [15:15] is separating and sorting things in [15:17] instead of common bins in the cities [15:20] these are placed at various places. So [15:22] we are literally giving an open API to [15:26] many parts of the world. We pretty much [15:28] the platform and the technology [15:30] provider. [15:32] As a byproduct when we were processing [15:35] billions of items, we realized [15:39] only few months ago that actually we can [15:42] also recognize who are behind these [15:45] materials. So only a month ago on 24th [15:49] June we launched a public facing [15:51] platform called deep nest to tell the [15:54] world of brands [15:56] that what really happens you know this [15:59] kind of information was never known [16:00] before 24th June every brand in America [16:04] every brand in Europe would exactly know [16:08] some are not happy about it would [16:11] exactly know what happens post bin story [16:15] you [16:16] Because many are known for greenwashing [16:18] many tell you many things many reports [16:21] are written even reports by United [16:24] Nations are written based on 500 samples [16:27] that's like calling a US election by [16:30] asking five people right that think [16:32] about 500 samples for [16:35] brands which produce hundreds of [16:37] billions of SKUs so this has become now [16:41] the world's largest data set where [16:42] millions of data points for every brand [16:45] is available for them to know how the [16:48] millions of dollars they're spending on [16:49] packaging [16:51] how it's actually going. You know, a [16:54] good example is Sprite is to be green. [16:58] Now it's become transparent white. If [17:01] those Sprite drinkers in the room, if [17:02] you would know that, this is the big [17:04] change. [17:06] Green sprite is PET. White transparent [17:11] Sprite is also PET. Just the color [17:14] difference, just by making it [17:16] transparent, it recovers 40% more. [17:20] That's a significant number. These [17:22] little things matters in the world of [17:24] recycling. It's not as simple as just [17:26] material sciences. Colors, sleeves, you [17:29] know, paper, caps, all factors matter. [17:34] And this kind of data wasn't available [17:36] before. [17:37] So just to conclude, you know, gray [17:41] parrot waste intelligence is making [17:44] things really actionable for multiple [17:48] stakeholders from a material recovery [17:50] facility point of view. It's bringing [17:53] them really more financial gains and [17:56] making [17:57] world of recycling more profitable. You [18:00] know they are already in it. So it's for [18:02] them it's a business. So the more [18:03] recoverable materials they are more they [18:06] can sell it in secondhand market. For [18:08] brands it's giving them packaging [18:10] intelligence they have huge impact on [18:12] supply chain and if they make the right [18:15] material choices they can it's better [18:17] for society is better for them and also [18:19] pressures from NOS's and Gen Z consumers [18:22] and others. We have a lot of information [18:26] uh on carbon credits generally because [18:28] we have information on pretty much every [18:30] commodity eventually becomes a waste. [18:32] And lastly, informing policy makers that [18:36] before you make changes to laws, it's [18:39] important that you make use of the data [18:41] available because there's very limited [18:44] insights on actually how recycling [18:47] works. [18:48] Thank you very much. [18:50] [Applause]