Ocado Runs 2,300 Warehouse Bots Under Single Centralized AI

Tom Scott

Because individual grid units lack autonomy and follow centrally computed paths, the automated facility functions as a single robot with thousands of physical actuators rather than a fleet of independent machines.

AI-assisted summary of How many robots does it take to run a grocery store? — original source. Check the original for full context.

Source published: 2021-07-05

Key ideas

Ocado's automated grocery warehouse operates across an area equivalent to seven football fields, routing over one million items daily using an aluminum storage grid up to 21 totes deep. (0:00)
Roughly 2,300 grid bots travel at speeds up to 14 km/h with clearances as tight as five millimeters, using cable hoists to retrieve bins containing 58,000 distinct product varieties. (0:42)
Predictive inventory placement allows incoming supplier stock to be shelved, picked, and loaded onto delivery vans in as little as five hours. (0:42)
The grid bots lack independent navigational autonomy and function under a centralized 'Hive Mind' AI, prompting the philosophical argument that the entire grid functions as a single robot with thousands of physical actuators. (4:06)

Detailed notes

## Background Published by Tom Scott on YouTube on July 5, 2021, this video visits an automated Ocado grocery warehouse to explore whether its thousands of moving mechanical boxes count as individual robots or a single giant hive mind. The independent, unsponsored piece was edited by Michelle Martin and developed from a title suggested by Grant Hurst.

## TL;DR Inside Ocado's massive automated grocery warehouse in south-east London, around 2,300 bots glide along an aluminum grid spanning the area of seven football fields, retrieving stacked bins with millimetric precision to process over a million items daily. While independent robotic packing arms handle delicate grocery items using 3D vision systems, the grid bots possess virtually no local decision-making authority. Instead, a central artificial intelligence called the Hive Mind dictates all paths, speeds, and logistics across the floor. Drawing parallels to termite colonies and the human gut microbiome, Tom Scott concludes that rather than thousands of autonomous machines, the entire operation is best understood as a single, distributed super-robot equipped with thousands of physical limbs.

## Atmosphere & style The video carries Tom Scott's characteristic intellectual curiosity, balancing awe at staggering industrial engineering with an accessible philosophical thought experiment. Fast-paced footage of thousands of metal boxes zipping past one another within hair's-breadth tolerances creates a mesmerizing, hypnotic sense of scale and precision. The narration moves seamlessly from concrete logistical statistics to biological metaphors, leaving the viewer with a lingering, mind-bending shift in how they conceptualize agency, identity, and robotic autonomy.

## Bulletin - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) Ocado's automated grocery warehouse operates across an area equivalent to seven football fields, routing over one million items daily using an aluminum storage grid up to 21 totes deep. - [00:42](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=42s) Roughly 2,300 grid bots travel at speeds up to 14 km/h with clearances as tight as five millimeters, using cable hoists to retrieve bins containing 58,000 distinct product varieties. - [00:42](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=42s) Predictive inventory placement allows incoming supplier stock to be shelved, picked, and loaded onto delivery vans in as little as five hours. - [04:06](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=246s) The grid bots lack independent navigational autonomy and function under a centralized 'Hive Mind' AI, prompting the philosophical argument that the entire grid functions as a single robot with thousands of physical actuators.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) The south-east London fulfillment warehouse replaces traditional human-oriented grocery aisles with an infrastructure engineered purely for automated machine throughput. - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) The facility spans seven football fields and processes over one million grocery items every day. - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) Grid bots operate with high precision, reaching speeds of 14 km/h while passing each other with margins as slim as five millimeters. - [00:42](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=42s) The Hive structure organizes 58,000 product lines into vertical columns of plastic totes stacked up to 21 bins deep. - [00:42](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=42s) Bots combine horizontal track movement with vertical cable hoists to extract specific totes, verifying their position via downward-facing optical sensors. - [00:42](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=42s) Predictive ordering algorithms reshuffle fast-moving products toward upper grid layers, enabling a five-hour turnaround from supplier delivery to customer dispatch. - [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) Robotic packing arms use twin 3D cameras and computer vision to calculate grip angles and suction forces for irregularly shaped consumer items. - [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) Automated packing arms independently enforce packing rules, such as preventing heavy canned goods from crushing delicate produce. - [02:49](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=169s) Unlike self-contained robotic arms, grid bots do not navigate autonomously; their routes, speeds, and destinations are orchestrated by a central 'Hive Mind' AI. - [02:49](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=169s) When bots report sensor discrepancies, the Hive Mind triggers an amber warning light and halts the unit so human perimeter operators can diagnose and command it wirelessly. - [04:06](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=246s) The warehouse architecture mirrors biological superorganisms like termite mounds or human microbiomes, where the boundary between individual units and a collective entity is ambiguous. - [04:06](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=246s) Because human intuition ties agency to physical scale and self-contained bodies, observers count thousands of bots when the grid is functionally one single robot with over 2,000 physical actuators.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) The Scale of the Hive

Inside an automated grocery packing warehouse in south-east London, customer orders are fulfilled without a single human walking up and down supermarket aisles. Instead of adapting machinery to spaces designed around human bodies, the entire facility—spanning an area equivalent to seven football fields—was custom-engineered from the ground up for maximum automated throughput. Across this massive expanse, thousands of mechanical rectangular units scurry along an elevated metal grid, moving over one million grocery items every single day.

```mermaid flowchart TD accTitle: Ocado Warehouse Automation Architecture accDescr: Flow diagram showing how customer orders move through the central Hive Mind, down through the grid bots and totes, into the automated packing arms, and out to delivery.

subgraph CentralControl["Central Intelligence"] HM["Hive Mind Central AI"] GO["Grid Operators (CCTV & Wireless Control)"] GO -.->|"intervenes on anomalies"| HM end

subgraph PhysicalGrid["The Hive Grid"] BOTS["~2,300 Grid Bots<br/>(14 km/h, 5 mm clearance)"] STACKS["Vertical Totes<br/>(Stacked up to 21 deep, 58,000 SKU types)"] HM -->|"movement profiles & waypoints"| BOTS BOTS -->|"cable hoist gripper"| STACKS end

subgraph PackingArea["Packing & Dispatch"] STATION["Pick Stations (Underneath Grid)"] ARMS["Robotic Packing Arms<br/>(3D Cameras & Vision)"] VAN["Delivery Vans<br/>(Out in ~5 hours)"] STACKS -->|"delivered by bots"| STATION STATION --> ARMS ARMS --> VAN end ```

The physical precision required to operate at this scale is striking. The boxes glide along tracks at speeds reaching roughly 14 kilometers per hour, missing each other by mere five-millimeter margins as they cross paths. Surveying the vast, mechanical grid stretching into the distance, Tom Scott notes that video cannot fully capture the sheer magnitude of the space, posing the central question of the operation: “So: how many robots are there in the Hive? Well, that depends on how you count them.”

- The facility covers the floor area of seven full-sized football fields. - Over 1,000,000 items are shuttled through the system daily. - Moving units travel at up to 14 km/h and operate with clearances as tight as 5 mm.

---

[00:42](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=42s) The Grid Architecture and Hoist Mechanics

The physical core of the warehouse is known as the Hive: an enormous, three-dimensional aluminum grid structure honeycombed with vertical storage columns. The products are stored inside standardized open plastic bins called totes, stacked directly on top of one another up to 21 bins deep. Operating across the top surface of this grid are approximately 2,300 bots tasked with sorting, retrieving, and shuffling inventory to fulfill incoming customer baskets.

Each bot maneuvers horizontally along the X and Y axes, but also operates along the vertical Z axis using an integrated cable hoist mechanism. When a specific item is required, a bot hoists down a mechanical gripper assembly into a grid column, latches onto the target tote, and lifts it into its internal bay. Navigation is handled locally by downward-facing optical sensors; every time a bot traverses a square, an underside laser registers the transition over the cell, constantly confirming its position grid square by grid square.

Managing 58,000 different product varieties requires aggressive predictive optimization. Using continuous streams of order data, the system forecasts purchasing trends and shifts fast-moving goods toward the upper layers and accessible perimeters of the grid. This logistical efficiency allows incoming stock unloaded from supplier lorries to be processed through the grid, packed into customer bags, and loaded onto delivery vans in as little as five hours.

- **Active fleet:** Approximately 2,300 bots operating concurrently. - **Storage capacity:** Stacks reach up to 21 totes deep within the grid columns. - **Product variety:** Around 58,000 distinct product types housed across the system. - **Turnaround speed:** New supplier inventory can arrive, be shelved, picked, and loaded into delivery vans in just five hours.

---

[01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) Automated Picking and Delicate Manipulation

Once the grid bots retrieve the appropriate totes, they transport them to stations situated directly beneath the aluminum framework. While human personnel still pack portions of the orders, a growing share of the handling is delegated to stationary robotic arms. Unlike moving rigid boxes on predefined tracks, handling groceries introduces serious mechanical hurdles because consumer packaging is engineered for human hands, resulting in unpredictable shapes, textures, and structural fragilities.

At each picking cell, twin overhead 3D cameras scan the interior of incoming storage boxes to build an image of the products inside. The arm's computer vision system identifies the item, confirms that it matches the live order requirements, calculates the optimal grasp angle and suction or grip force, and lifts the item into the customer's delivery bag. The algorithm must continuously weigh physical constraints—such as ensuring heavy canned goods are never placed on top of soft baked goods or produce. Items that are extremely heavy, awkwardly large, or unusually fragile remain off-limits to the automated arms, leaving a broad spectrum of intermediate everyday products for the machines to pack safely in real time.

---

[02:49](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=169s) The Hive Mind vs. Distributed Intelligence

The contrast between the packing arms and the grid boxes highlights a fundamental division in robotics architecture. A packing arm clearly fits the textbook definition of an individual robot: it has an isolated controller, analyzes visual sensor feeds independently, and makes autonomous real-time decisions about physical interactions. By contrast, the bots scurrying across the top of the grid operate on an entirely different paradigm governed by a central artificial intelligence dubbed the "Hive Mind."

The Hive Mind exercises absolute spatial awareness over the facility, tracking the location and vector of every tote and machine from the moment goods enter the building. While individual bots run onboard motion-profile software to handle motor acceleration, wheel braking, and precision stops, their destinations, routes, and timing are dictated entirely by the central AI. When an anomaly occurs—such as an onboard sensor reporting unexpected data back to the central server—the Hive Mind flags the discrepancy, illuminates an amber warning light on the unit's chassis, and brings it to an immediate controlled stop. Grid operators stationed along the warehouse perimeter monitor these events via an array of overhead CCTV cameras, diagnosing errors and broadcasting wireless commands directly to the bot to resolve track deadlocks remotely.

---

[04:06](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=246s) Superorganisms, Microbiomes, and the Nature of Identity

The centralized command structure of the warehouse blurs conventional distinctions between a collection of individual machines and a single unified organism. Tom points out that biology frequently struggles with this exact philosophical boundary: a termite mound can be viewed either as millions of separate insects or as a singular, distributed superorganism. Similarly, scientific discoveries over the past two decades have shown that human gut bacteria synthesize and metabolize neurotransmitters that directly influence human brain chemistry and mood, yet few people intuitively consider their digestive microbiome to be an intrinsic part of their personal identity.

Humans habitually assign agency, identity, and intelligence to entities that match human physical scale and speed. Consequently, watching thousands of distinct wheeled boxes zip across an aluminum frame naturally prompts observers to count several thousand distinct robotic entities. However, given that none of the boxes possess meaningful autonomy or independent goals—serving merely as remote physical actuators executing paths assigned by a single central computational engine—Tom argues that our intuitive count is mistaken: “I'd say there's a good argument that that is one robot... with 2,000 ways to interact with the world.”

## Mentioned Casual Historian, Grant Hurst, Michelle Martin, Ocado, P. Strandwitz, The Hive, Tom Scott

## Video description In Ocado's grocery warehouses, thousands of mechanical boxes move on the Hive. Are they all individual robots? Or is this one giant hive mind? • Thanks to Ocado: https://www.ocadogroup.com/solutions/online-grocery (this video is not sponsored, and they had no editorial control).

Reference: Strandwitz P. (2018). Neurotransmitter modulation by the gut microbiota. Brain research, 1693(Pt B), 128–133. https://doi.org/10.1016/j.brainres.2018.03.015 -- I should probably have put a "might" in there, but that's a thorough review of research and I think the claim just about holds up!

Edited by Michelle Martin https://www.youtube.com/@OnTheCrux

Filmed safely: https://www.tomscott.com/safe/

Thanks to Grant Hurst from Casual Historian for suggesting the title: https://www.youtube.com/user/grantghurst

🟥 MORE FROM TOM: https://www.tomscott.com/ (you can find contact details and social links there too) ❓ LATERAL, free weekly podcast: https://lateralcast.com/ https://youtube.com/lateralcast/ 👥 THE TECHNICAL DIFFICULTIES: https://youtube.com/techdif

[Source](https://www.youtube.com/watch?v=ssZ_8cqfBlE)

中文简报

Ocado依托单一中心化AI调度2300台仓储机器人由于单个网格机器人缺乏自主权并遵从集中计算的路径，该自动化设施的运作形态更像是一个拥有数千个物理执行器的单体机器人，而非一支由独立机器组成的机队。在Ocado位于伦敦东南部的巨型自动化食品杂货仓库内，约2300台机器人在占地相当于七个足球场的铝制网格上滑行，以毫米级的精度调取堆叠的料框，每日处理超过一百万件商品。尽管独立的机械包装臂利用3D视觉系统处理易碎的杂货，但网格机器人本身几乎没有任何本地决策权。相反，名为“蜂巢思维”（Hive Mind）的中央人工智能统筹着整个作业区域的所有路径、速度和物流调度。类比于白蚁群体和人体肠道微生物菌群，Tom Scott总结认为，与其将整个运营系统视为数千台自主机器，不如将其理解为一个配备了数千个物理肢体的单一分布式超级机器人。 Ocado的自动化食品杂货仓库运作范围相当于七个足球场大小，利用深度达21层料框的铝制存储网格，每日分流超过一百万件商品。 (0:00) 约2300台网格机器人以高达14公里/小时的速度穿梭，间距紧凑至仅五毫米，并通过钢缆起重系统调取存放着58000种不同品类商品的料框。 (0:42) 预测性库存布局使供应商送达的货物从上架、拣选到装入配送货车的最短耗时仅需五小时。 (0:42) 网格机器人缺乏独立的导航自主权，完全在中心化“蜂巢思维”AI的控制下运行，从而引发了一种哲学层面的探讨，即整个网格本质上是一个拥有数千个物理执行器的单体机器人。 (4:06)

Reading summary

Inside an automated grocery warehouse in south-east London, thousands of mechanical units fulfill orders across a massive grid without human pickers walking aisles. Presenter Tom Scott argues that because these units lack independent autonomy and answer to a single central intelligence, the facility is better understood as a single massive robot rather than thousands of separate machines. High-Density Grid Architecture and Rapid Retrieval (0:42) Spanning an area equivalent to seven football fields, the warehouse is organized around the Hive, an aluminum grid honeycombed with vertical storage columns holding standardized plastic totes stacked up to 21 bins deep. Operating across the surface, approximately 2,300 bots travel at speeds up to 14 kilometers per hour with clearances as narrow as five millimeters. Downward-facing optical sensors track movement grid square by grid square, and onboard cable-hoist grippers lower down into the columns to retrieve and lift targeted totes. To manage roughly 58,000 product varieties and move over one million items daily, predictive algorithms analyze order patterns and continuously relocate high-velocity stock toward the grid's upper layers and edges. This spatial optimization allows newly unloaded supplier inventory to be ingested, retrieved, packed, and loaded onto delivery vans in as little as five hours. Robotic Picking for Fragile Consumer Packaging (1:55) Once retrieved by grid bots, totes are lowered to picking stations where stationary robotic arms handle an increasing share of item packing alongside human workers. Handling commercial groceries poses distinct challenges because packaging is manufactured for human hands rather than rigid automated grippers, resulting in varied shapes, textures, and structural fragilities. At each station, twin overhead 3D cameras scan incoming totes so the arm's computer vision system can identify items, match them to live orders, and calculate grasp angles and grip forces. The control algorithms enforce physical placement constraints, such as ensuring heavy canned goods are never stacked on top of soft produce or baked goods. Items that are unusually heavy, large, or fragile remain outside the automated arms' parameters and continue to be packed manually. Centralized AI Navigation versus Isolated Controllers (2:49) The facility demonstrates two fundamentally different approaches to robotics. While stationary picking arms function as conventional autonomous robots—relying on isolated controllers and onboard vision systems to make independent physical decisions—the grid bots operate entirely under a centralized artificial intelligence known as the "Hive Mind." The Hive Mind maintains global spatial awareness across the facility, coordinating every tote and determining the specific trajectories, routes, and timings for every bot. Individual units run local software solely for acceleration, braking, and stopping. When an onboard sensor detects an anomaly, the central AI halts the bot and illuminates a warning indicator, prompting perimeter human operators to inspect the situation via CCTV and transmit wireless override commands to clear track deadlocks. The Superorganism Paradigm and Defining a Robot (4:06) The central control model blurs the philosophical boundary between separate machines and a single integrated system. Tom Scott draws a comparison to biological superorganisms, noting that science struggles with similar classifications when viewing termite colonies as individual insects versus single entities, or considering human gut microbiomes that regulate brain chemistry without being perceived as part of human identity. Scott observes that humans intuitively assign individual identity to physical objects that match human scale and motion, leading visitors to count thousands of distinct robots zipping across the grid. However, because none of the individual units possess autonomous goals, he concludes that the grid is more accurately categorized as a single computational robot operating with thousands of remote physical actuators. The facility illustrates how purpose-built automated infrastructure can optimize retail distribution at massive scale, while challenging conventional assumptions about machine agency by demonstrating that thousands of coordinated moving parts can function conceptually as a single machine.

精炼阅读摘要

Tom Scott 在探访位于伦敦东南部的大型自动化食品杂货配送中心后指出，尽管仓库表面上由数千台穿梭小车协同运作，但在系统控制与计算架构的本质层面上，该设施更应被视作一个由中央人工智能统一驱动、拥有数千个远程物理执行器的单一超级机器人。高密度网格结构与立体物流调度 (0:00) 该配送中心占地面积相当于七个标准足球场，专为纯自动化作业量身定制，内部不存在供人工行走的传统货架通道。系统核心是一个被称为“蜂巢”（The Hive）的三维铝制立体网格结构，内部垂直堆叠了深达 21 层的标准化塑料料箱，容纳约 58,000 种不同品类的商品。每日有超过一百万件杂货在此流转。约 2,300 台方形机器人在网格顶表面以高达 14 公里/小时的速度穿梭，交会时彼此间距仅为 5 毫米，并通过底部激光传感器实时扫描网格方格确认自身坐标。为了兼顾存取效率，每台小车配备了缆绳起重夹爪机构，能在 X 轴与 Y 轴平移到位后沿 Z 轴向下垂降夹具，将目标料箱提升至内部货舱。整个系统依赖连续订单流进行预测性优化，持续将高频次采购的热门商品调配至网格表层或易于存取的边缘区域。这种高效调度使供应商卡车卸下货物到重新装配成客户订单并装载至配送货车的最快周期缩短至约五小时。末端分拣视觉与物理抓取约束 (1:55) 从网格中取出的料箱会被输送至下方的拣选工作站。虽然部分订单仍由人工打包，但固定式工业机械臂承担的比例正在增加。杂货处理的难点在于消费品包装最初是为人类双手设计的，存在不可预测的外形、材质表面和易损结构，无法直接采用简单的固定轨道搬运逻辑。每个拣选工位配备两台顶置 3D 相机，用于扫描料箱内部并建立三维图像。机械臂的计算机视觉系统会自动核对实时订单信息，计算最优抓取角度、夹持力或吸盘吸力，将商品装入配送袋。控制算法必须时刻计算物理负荷约束，例如绝对不能将沉重的罐头压在柔软的面包或生鲜果蔬之上。目前，极重、异形或极其易碎的商品仍被排除在自动分拣之外，机械臂主要负责处理处于中间区间的各类日常标准包装商品。 “蜂群思维”中央智能与执行器分工 (2:49) 拣选机械臂与网格搬运小车代表了两种截然不同的机器人架构范式。拣选机械臂拥有独立的控制器和视觉分析单元，能自主做出物理决策，符合传统意义上的单体机器人定义；而网格上的搬运小车则完全受控于名为“蜂群思维”（Hive Mind）的中央人工智能控制系统。蜂群思维对全场保有全局空间感知，自货物入库起便实时追踪每个料箱与机器人的位置与运动向量。个别小车搭载的车载运动软件仅负责执行加速度计算、车轮制动和精准停靠等基础动力控制，其行进路线、任务端点和移动时序均由中央 AI 强制下达。一旦小车传感器报告异常，中央系统会立即触发预警并亮起黄色警示灯让其受控停机，随后由仓库周边的调度人员通过闭路电视网络研判故障，并无线发送指令以远程解除轨道拥堵。超级有机体隐喻与机器自主性重思 (4:06) 这种高度集中的控制架构打破了传统上将单个机械硬件视为独立个体的认知习惯。Tom Scott 引入生物学现象进行类比：白蚁丘既可被看作数百万只独立昆虫的聚集，也可被视为一个统一运作的超级有机体；人类肠道微生物群虽能合成代谢神经递质并直接影响大脑情绪，但极少有人将其视作自身人格认同的组成部分。人类本能上倾向于将智能与自主实体投射到那些体积、运动速度与人类尺度相仿的独立物体上。然而，由于网格中的两千多台小车本身既无自主规划能力也无独立目标，仅是作为执行中央计算引擎路径规划的远程物理执行端，Tom Scott 认为更合理的界定方式是将其视为一个“拥有 2,000 种与物理世界交互方式的单一机器人”。该系统表明，超大规模自动化仓储的技术前沿已不再局限于单机设备的机械灵活性，而是转向全局集中算法对海量物理执行终端的高效统辖。这一架构不仅重塑了物流供应链的时效极限，也对工程设计与哲学层面关于“何为独立机器人”的传统定义提出了挑战。

Detailed notes — zh

## Background 该视频由知名科技与科普创作者汤姆·斯科特（Tom Scott）于2021年7月5日发布在YouTube平台上。视频简介探讨了在Ocado的生鲜食品仓库“蜂巢”（The Hive）中穿梭的数千个机械箱子，究竟是各自独立的机器人，还是属于一个庞大“蜂群思维”控制下的超级单一整体。 ## TL;DR 视频深入探访了Ocado位于伦敦东南部、面积达七个足球场的全自动化智能生鲜仓库。在这个被称为“蜂巢”的金属网格结构中，约2,300个机器箱子以高达14公里的时速和仅5毫米的间距高速协同作业，负责在深达21层的网格中调取并搬运58,000种商品。随后，配备3D相机的独立机械臂接管包装环节，根据物品材质与重量实时规划抓取与打包逻辑。全场由一个强大的中央人工智能“蜂群思维”统一调度与纠错，从而引出了关于自动化系统中“个体机器人”与“单一庞大超个体”的界定与哲学思考。 ## Atmosphere & style 全片节奏轻快明朗，延续了汤姆·斯科特标志性的沉浸式科普叙事风格，带领观众置身于震撼且略带超现实感的非人尺度工业奇观之中。现场成百上千台机械箱子静谧而高速穿梭的画面，既展现出精密工程学带来的工业美感，又透露着一种秩序井然的科幻感。随着叙事从具象的物流运作层层递进到对意识、群体与机器定义的哲学探讨，观众的情绪也从最初对庞大阵列的纯粹惊叹，自然转化为对自动化本质的深度玩味与沉思。 ## Bulletin - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) Ocado在伦敦东南部运营着相当于7个足球场大小的巨型智能仓库，依靠网格系统每日可分拣处理超过100万件商品。 - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) 仓库网格上运行着约2,300个机器人，最高时速达14公里，在以深达21层的周转箱阵列上高速交汇时安全间距仅有5毫米。 - [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) 包装站引入配备双3D相机的自主机械臂实时抓取打包，同时遵守“重物不压软物”等复杂物理堆叠规则。 - [03:01](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=181s) 全场机器受中央人工智能“蜂群思维”统一路径规划，引发了其究竟是2,000个独立机器人还是单一“超个体”控制器的哲学探讨。 ## Key takeaways - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) Ocado的智能仓库“蜂巢”（The Hive）完全抛弃传统超市格局，专为极致机器运作效率定制非人尺度空间。 - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) 仓库存储约58,000种单品，商品堆叠在塑料周转箱内，最深可达21层。 - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) 网格机器人具备三维操作能力，能在水平X/Y轴快速移动，并依靠钢索起重系统与抓取装置深入垂直Z轴抓取货物。 - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) 每个机器人底部均配备激光传感器，通过网格单元格时持续校准确认位置以保证极高精度。 - [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) Ocado结合大数据预测消费者购买习惯，将热门高频商品放置在网格最上层，实现到库商品5小时内快速出库。 - [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) 传统食品包装专为人类手部与感官设计，机器人物理抓取面临材质脆弱与形状各异的巨大挑战。 - [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) 包装站上方的双3D相机在机械臂操作前拍照识别最佳抓取点，并验证货品正确性及避免压损柔软商品。 - [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) 与网格小车不同，包装机械臂拥有独立控制器与传感器自主决策系统，具备明确的个体机器人属性。 - [03:01](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=181s) 整个网格受中央AI“蜂群思维”统一调配，机器人负责精准执行动作曲线，异常时会自动亮橙灯急停等待远程修复。 - [03:01](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=181s) 如同白蚁群或肠道菌群与人体的关系，高度中心化控制的网格系统在哲学层面上更像是一个拥有2,000个行动终端的单一巨型机器人。 ## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=0s) 巨型“蜂巢”：超越人类尺度的智能仓库在伦敦东南部，坐落着一座占地面积相当于七个足球场的巨型网上超市配货仓库。这里没有传统超市里推着购物车在货架间穿梭的顾客或理货员，取而代之的是一个被称为“蜂巢”（The Hive）的庞大金属网格结构。在这个完全为了追求极致效率而定制的非人尺度空间里，成百上千个像小箱子一样的机器人每天在网格顶部穿梭，日均处理超过100万件商品。主持人汤姆·斯科特（Tom Scott）站在现场感叹，摄像机镜头根本无法完整呈现这个地方究竟有多么令人震撼地巨大。这些机器箱子的最高时速可达14公里，而在高速移动时，它们彼此之间的距离仅仅只有5毫米。这个庞大的系统究竟是如何运作的？据现场工作人员介绍，金属网格结构中整齐地叠放着装满商品的塑料周转箱（totes），部分区域的堆叠深度甚至达到了21层。在这个网格之上，运行着大约2,300个机器人（bots）。这些机器人不仅能在X轴和Y轴（水平方向）上移动，还能通过钢索起重系统（cable hoist system）将抓取装置（gripper assembly）向下延伸到Z轴（垂直方向），从而精准地抓取网格深处的周转箱。每个机器人的底部都配备了激光传感器，每当它们经过网格上的一个单元格时，传感器就会记录并确认一次位置。整个网格中存储了大约58,000种不同类型的单品。Ocado利用收集到的海量数据来预测客户最可能购买的商品，并将这些热门商品放置在机器人最容易触及的网格上层。正如工作人员所说： > “所有这些因素结合在一起，意味着我们可以在短短五个小时内，将供应商送达仓库的货物重新分拣并装入送货车中。” ```mermaid flowchart TD accTitle: Ocado智能仓库运作流程 accDescr: 展示从商品入库、在网格中存储、由蜂群思维调度、机械臂/人工打包，到最后装车配送的完整流程。 subgraph Grid["网格存储区 (The Hive)"] A["商品入库"] --> B["放入周转箱 (Totes)"] B --> C["2300个网格机器人 (Bots)"] C -->|"X/Y/Z轴移动与抓取"| D["深达21层的网格阵列"] end subgraph Control["控制与优化"] E["蜂群思维 (Hive Mind AI)"] E -->|"路径规划与调度"| C F["历史数据分析"] -->|"预测需求/优化布局"| E end subgraph Packing["包装与出库"] D -->|"运送至包装站"| G["包装区域"] G -->|"3D相机定位"| H["机械臂/人工打包"] H -->|"避免重物压软物"| I["装入配送车 (5小时内出库)"] end ``` - **仓库规模**：相当于7个足球场大小。 - **日处理量**：超过100万件商品。 - **机器人数量**：约2,300个。 - **移动精度**：最高时速14公里，并排通过时仅相距5毫米。 - **存储深度**：部分网格区域深达21个周转箱。 - **商品种类**：约58,000种。 - **周转时间**：商品从入库到装上配送车最快仅需5小时。 --- ### [01:55](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=115s) 机械臂的挑战：如何像人类一样打包商品当网格机器人将装有商品的周转箱运送到边缘后，这些箱子会被送往包装站。在这里，部分打包工作仍由人工完成，但越来越多的工作正逐步交由机械臂来处理。然而，让机器打包面临着一个巨大的挑战：市面上所有的食品包装都是为了人类的双手和感官而设计的，不同物品之间的物理交互和材质差异极其复杂。在网格结构下方的包装站里，机械臂正在对活生生的客户订单进行实时抓取。每个包装站上方都安装了两个3D相机，在机械臂抓取前，相机会对周转箱内部进行拍照，分析并决定最佳的抓取点，同时确认商品是否正确。接着，机械臂会将商品放入客户的购物袋中。在打包过程中，系统有着严格的物理逻辑，例如绝对不能把重物（如罐头）压在柔软的商品（如面包）上面。虽然目前机械臂还无法处理过于沉重、体积过大或极度易碎的物品，但它们已经能够轻松应对绝大多数常规商品。对于这些包装机械臂，我们很容易将其定义为“独立的机器人”。因为每一个机械臂都拥有自己独立的控制器，并且必须依靠自身的相机和传感器，在实时状态下做出大量的抓取决策。 > “一旦货物被送达，它们就归该机器人控制，由它来决定如何抓取，随后网格机器人会重新取走周转箱。” --- ### [03:01](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=181s) 蜂群思维：是两千个独立个体，还是一个超级大脑？然而，相比于拥有独立控制器的包装机械臂，网格上那2,300个移动的机器箱子，其身份界限就变得模糊得多了。控制这整个庞大网格运转的，是一个被称为“蜂群思维”（Hive Mind）的中央人工智能系统。从商品进入仓库的那一刻起，这个“超级大脑”就会追踪它的每一次移动。 “蜂群思维”会精确地告诉每一个机器人应该在什么地方停下，而机器人自身的控制系统则负责执行具体的运动曲线，确保自己能严丝合缝地停在指定位置。由于机器人之间的安全间距只有5毫米，任何微小的偏差都可能导致碰撞。如果某个机器人的传感器检测到的数据与“蜂群思维”的预期不符，它的指示灯就会变成橙色，并自动安全停下以接受检查。此时，坐在网格边缘的操作员会通过上方的CCTV监控系统观察该机器人的状况，并通过无线网络向其发送指令进行远程修复。这种高度集中的控制模式引发了一个哲学上的思考：我们该如何定义“个体”与“群体”？这就像白蚁丘是一个由数百万只白蚁组成的合作社，还是一个单一的“超个体”？又比如人类肠道中的菌群，它们不仅在消化系统里生存，还会产生和消耗影响人类思维的神经递质，但很少有人会把肠道细菌视为“自己”的一部分。人类习惯于将智慧和身份归诸于那些与我们体量相当的实体。因此，当我们看着网格上2,000多个移动的箱子时，很容易脱口而出“这里有2,000个机器人”。但汤姆·斯科特给出了不同的看法： > “但由于只有一个控制器，且这些箱子只是在遵循分配给它们的路径，我认为完全有理由说，这就是一个机器人……只是它有2000种与世界互动的方式。” ## Notable quotes > [03:01](https://www.youtube.com/watch?v=ssZ_8cqfBlE&t=181s) 由于只有一个控制器，且这些箱子只是在遵循分配给它们的路径，我认为完全有理由说，这就是一个机器人……只是它有2000种与世界互动的方式。 — Tom Scott ## Mentioned Tom Scott, Ocado, 蜂巢 (The Hive), 蜂群思维 (Hive Mind), Michelle Martin, Grant Hurst, Peter Strandwitz ## Video description In Ocado's grocery warehouses, thousands of mechanical boxes move on the Hive. Are they all individual robots? Or is this one giant hive mind? • Thanks to Ocado: https://www.ocadogroup.com/solutions/online-grocery (this video is not sponsored, and they had no editorial control). Reference: Strandwitz P. (2018). Neurotransmitter modulation by the gut microbiota. Brain research, 1693(Pt B), 128–133. https://doi.org/10.1016/j.brainres.2018.03.015 -- I should probably have put a "might" in there, but that's a thorough review of research and I think the claim just about holds up! Edited by Michelle Martin https://www.youtube.com/@OnTheCrux Filmed safely: https://www.tomscott.com/safe/ Thanks to Grant Hurst from Casual Historian for suggesting the title: https://www.youtube.com/user/grantghurst 🟥 MORE FROM TOM: https://www.tomscott.com/ (you can find contact details and social links there too) ❓ LATERAL, free weekly podcast: https://lateralcast.com/ https://youtube.com/lateralcast/ 👥 THE TECHNICAL DIFFICULTIES: https://youtube.com/techdif [Source](https://www.youtube.com/watch?v=ssZ_8cqfBlE)

Source quotes

[3:01] Tom Scott: 由于只有一个控制器，且这些箱子只是在遵循分配给它们的路径，我认为完全有理由说，这就是一个机器人……只是它有2000种与世界互动的方式。

Original transcript

[0:00] - This is a grocery packing warehouse in the south-east of London. [0:03] Folks have ordered their groceries online, and instead of people walking [0:05] up and down the aisles of a traditional grocery store to pick up items, [0:09] these little devices move more than a million items every day between them. [0:13] Instead of trying to build machines that fit into a world designed for humans, [0:16] this entire facility, the size of seven football fields, [0:20] is custom-built to make this as efficient as possible. [0:22] I cannot get across on camera [0:25] just how incredibly large this whole place is. [0:29] The top speed of those boxes is about 14 kilometres an hour [0:32] and they pass within five millimetres of each other when they move. [0:35] So: how many robots are there in the Hive? [0:39] Well, that depends on how you count them. [0:42] - What you're looking at behind me is the Hive. [0:44] We've got the grid, which is the metal structure, [0:48] it's filled with products, [0:49] so that product is housed within what we call totes [0:52] or these open boxes that you can see at the top of the grid. [0:56] We have around 2,300 bots [0:58] and they move the stock around the warehouse, [1:00] so that we can pair it up with customer orders to be delivered. [1:04] The bots will move in the X and Y direction, [1:06] they can also move in the Z direction. [1:08] So we have a cable hoist system. [1:10] That means that they can lower down a gripper assembly [1:13] and grab onto the stock that we have in the grid. [1:16] In some areas we have as deep as 21 totes, all in a line. [1:21] We have sensors underneath the bot, [1:23] that mean that every time it passes over one of the squares on the grid, [1:27] it has a laser underneath it, [1:28] that can tell when it moves over that cell. [1:30] So it will say "I passed a cell, I passed a cell, I passed a cell." [1:34] We have around 58,000 different types of individual products within the grid. [1:39] We'll use the huge amounts of data that we gather [1:41] to understand what customers are most likely to order [1:44] and place them in the easiest to reach locations for bots. [1:48] All of that combined means that we can have stock [1:50] into the warehouse and out again from our suppliers, [1:53] into a delivery van, in just five hours. [1:55] - The robots deliver their crates over to the packing area. [1:58] Some of that packing is done by humans, but there are mechanical arms as well. [2:02] One of the big problems they've got to solve [2:04] is that grocery packaging is designed for people, [2:06] and the interactions between all the objects [2:09] they deal with can be... complicated. [2:10] - Currently, we're standing underneath the grid structure [2:13] that we saw earlier. The robots that we saw [2:15] are actually dropping off products to the station. [2:18] Once they're dropped off, they become in control of the robot [2:20] and the robot decides how to pick them, [2:22] before the grid robot picks up their tote and takes it away again. [2:25] There's two 3D cameras above the totes, [2:27] so it takes an image of the storage box before it picks from it [2:31] decides on the best grasp point, [2:32] confirm that the product is correct, [2:34] and then place it into a customer's shopping [2:36] and it's doing all of that on live customer orders. [2:38] We make sure that we don't mix something that was heavy [2:40] with something that's soft. [2:42] There are things that the robot can't pack, [2:43] things such as very, very heavy, very large or very fragile items, [2:47] but there's a lot of range in between that. [2:49] - It's fairly safe to call each individual packing arm "a robot". [2:52] Each one has a separate controller [2:54] and it has to make a lot of decisions in real time [2:56] using cameras and sensors on each individual arm. [2:58] But here in the Hive, it's not quite as simple. [3:01] - Controlling all of this is what we refer to as the Hive Mind, [3:05] so this artificial intelligence system that controls everything. [3:09] Once the stock comes in, [3:11] from that exact point that it enters the system, [3:14] the Hive Mind will track its every single movement. [3:16] The Hive Mind will tell the bots exactly where they need to stop, [3:20] but it is the bot's control system [3:22] that allows it to follow a movement profile [3:25] and stop exactly as the hive mind expects it to. [3:27] There's just a 5mm gap between bots, [3:30] so that's both as they pass each other and if they stop next to a bot as well. [3:35] This bot with the orange light [3:37] may have experienced something within its sensors [3:39] and all of those readings it's taking, [3:41] that when it feeds that back to the Hive Mind, [3:44] it's not quite what the Hive Mind was expecting, [3:47] so it just brings it to a safe stop for investigation. [3:50] We do have a number of grid operators that sit at the edge of the grid [3:55] and they're able to use our CCTV system that we have above the grid [3:59] to look at exactly what's going on with that bot [4:02] and they can try and fix that remotely and send commands to the bot wirelessly. [4:06] - The boundary between "individual" and "group" is a little blurry. [4:09] Is a termite colony a collection of individuals working together, [4:12] or one big superorganism? [4:14] Is the microbiome in your gut part of you? [4:16] Because in the last couple of decades scientists have started to work out [4:19] that the bacteria in your digestive system produce and consume neurotransmitters [4:23] that affect how you think. [4:24] But most people wouldn't consider their gut bacteria to be "part of them". [4:29] We tend to ascribe intelligence and identity [4:31] to things that act at roughly our scale. [4:33] It's easy to look at the Hive, [4:34] see the moving boxes and go, "okay, 2000 robots". [4:37] But with only one controller [4:39] and the boxes just following the paths they're given, [4:42] I'd say there's a good argument that [4:43] that is one robot... [4:45] with 2,000 ways to interact with the world. [4:48] Thank you so much to everyone here at Ocado. [4:50] There is a link to them [4:51] and more about their technology in the description.