Chaos and Order

💡 왼쪽 원문을 읽으면서 오른쪽에 따라 써보세요. Tab 키로 힌트를 받을 수 있습니다.

원문 렌더가 준비되기 전까지 텍스트 가이드로 표시합니다.

Introduction — In May 2026, Warehouse Automation Is No Longer "Vision"

As recently as 2022, warehouse robotics was a "story that only fits operators like Amazon." The landscape in May 2026 is unrecognizable. **E-commerce volume growth + chronic logistics labor shortages + GPU-based simulation learning** combined to push robots into everyone from mid-tier 3PLs to global retailers. The market is no longer a single solution but a **layered stack**.

This article is not marketing copy. It is an honest read of "what slots into which seat on the floor today." AMR/AGV, Goods-to-Person shuttles, piece-picking arms, Amazon Robotics, humanoid pilots, the NVIDIA Isaac learning stack, and Korean/Japanese local players — all compared with real 2026 specs.

Warehouse Robotics 2026 — Decomposing the Stack Into 7 Layers

Start with the big picture. The standard warehouse robotics stack in 2026 splits into seven layers.

1. **Mobile (AMR/AGV)**: Autonomous robots that work alongside humans.

2. **Goods-to-Person (G2P) shuttles/cubes**: Bring shelves/totes to the operator.

3. **Piece/case picking arms**: Vision + gripper to grab individual SKUs.

4. **Fixed conveyor + sortation**: Classifies and routes.

5. **Humanoid/bipedal**: Human-friendly form factor, pilot stage.

6. **Software (WMS + WES + simulation)**: Orchestration and learning.

7. **Last-mile + drone delivery**: Outside the warehouse, to the customer doorstep.

Each layer is a separate market, and the adoption order varies by company. We break them down one by one below.

AMR vs AGV — The Gap Decisively Widened in 2026

AGVs (Automated Guided Vehicles) are first-generation automation that follow magnetic strips or QR codes embedded in the floor. As of 2026, more than 90% of new deployments switched to **AMRs (Autonomous Mobile Robots)**, the second generation driven by LiDAR + cameras + SLAM. The difference is simple.

- **AGV**: Static paths, infrastructure work required, must be separated from humans.

- **AMR**: Dynamic paths, no infrastructure needed, shares aisles with humans.

The representative AMR/AGV players are as follows.

- **Locus Robotics — Locus Origin**: U.S. The de facto standard in "collaborative picking." Global partnerships with DHL/GXO.

- **6 River Systems**: Acquired by Shopify in 2019, then sold to Ocado in July 2023. Now part of Ocado Intelligent Automation.

- **GreyOrange Ranger**: India-born, globally expanding. The GreyMatter multi-agent orchestration is its core.

- **Geek+**: China-born, EQT invested in 2024. Strong across APAC + Europe.

- **Hai Robotics**: The pioneer of case-handling + tote-handling ACR (Autonomous Case-handling Robot).

- **Quicktron**: China-born, deeply integrated with Cainiao (Alibaba logistics).

- **Fetch Robotics**: Acquired by Zebra Technologies in 2021, integrated into the Symmetric Vision line.

- **Vecna Robotics**: U.S., focused on large-payload autonomous forklifts and pallet jacks.

- **inVia Robotics**: U.S., targeting mid-tier warehouses with a SaaS model ("Robotics-as-a-Service").

- **MiR — Mobile Industrial Robots**: Denmark-born, Teradyne subsidiary. Industrial, healthcare, retail back-of-house, and more.

- **OTTO Motors**: Canada-born, acquired by Rockwell Automation in October 2023. Strong in heavy-payload AMRs.

A point that often gets missed in adoption decisions: **"Is our warehouse in pick mode or move mode?"** Models like Locus where the picker walks alongside, and models like OTTO that focus on unmanned transport, have completely different ROI curves.

Goods-to-Person (G2P) — The Era of Shuttles and Cubes

If AMRs "follow the human," G2P "brings the shelf to the human." Storage density is high and workstation throughput is consistent, which is why companies maximizing throughput-per-square-meter choose it. Representative players are as follows.

- **AutoStore**: Norway-born. A 3D cube grid + top-running robots that pull totes up. The de facto standard among global retailers.

- **Symbotic**: U.S.-born, multi-site contract with Walmart. June 2022 IPO brought it to roughly USD 11 billion market cap. The apex of multi-agent case handling.

- **Exotec — Skypod**: France-born, 3D grid + proprietary shuttles. Adopted by Uniqlo/Decathlon/Gap.

- **Hai Robotics**: Beyond ACR, also has G2P case/tote systems.

- **Swisslog AutoStore + CarryPick**: KUKA subsidiary. AutoStore integration + proprietary mobile shuttles.

- **Knapp**: Austria-born. OSR Shuttle Evo + Pick-it-Easy.

- **TGW Logistics**: Austria-born. FlashPick + Stingray shuttles.

Selection is not simple. **AutoStore is tote-unit, Symbotic is case-unit, Exotec sits between** — storage units differ. SKU turnover, average order lines, and storage unit volume are the biggest variables.

Berkshire Grey + SoftBank — After the Acquisition

Berkshire Grey is a U.S. company founded by former iRobot CTO Tom Wagner in 2013. It differentiated with "integrated AI fulfillment" that bundles mobile + piece picking + sortation. After going public via SPAC in 2021, the stock languished.

SoftBank announced acquisition in June 2023, and the **deal closed in April 2024**, taking the company private. Intent was strong: tie together the big picture inside SoftBank Vision Fund's portfolio that includes Symbotic and AutoStore. The core post-merger products are as follows.

- **BG Flex AI**: Mobile picking cells with AMRs + picking arms integrated.

- **BG Sort with AI**: AI-vision-based sorter, recognizes SKUs without labels.

- **BG Picking**: Cellular piece picking, dual vacuum + finger grippers.

Since joining SoftBank's portfolio, sales activity in Japan and Korea has re-energized.

Piece Picking — Covariant Joins Amazon and the Category Reshapes

Piece picking (grabbing a single SKU and moving it elsewhere) was the "hardest remaining problem" through the early 2020s. After 2024, **foundation models + multimodal learning** delivered a clear step change.

- **Covariant**: U.S.-born. Famous for its multi-task transformer-style "RFM (Robotics Foundation Model)." In August 2024, Amazon acquired the three core founders + key engineers + license (a "reverse acqui-hire"), effectively absorbing it. The Covariant corporate entity remains.

- **RightHand Robotics**: U.S.-born. RightPick series. ABB acquired it in November 2024 and integrated it into ABB Robotics.

- **Pickle Robot**: U.S.-born. Specializes in truck unloading, with a form factor that lifts one box at a time.

- **Plus One Robotics**: U.S.-born. Yonder, a cloud-supervised control plane, is its differentiator.

- **Mujin**: Japan-born. The deterministic motion planner "MujinController," PLC-tier in capability, is the core. Backed by Sequoia and Toyota AI Ventures.

- **Soft Robotics**: U.S.-born. Vacuum + soft-finger dual grippers. Onto Innovation acquired the core IP in 2023.

Piece-picking ROI is decided by **SKU diversity and box variety**. Single-SKU lines favor deterministic motion, while tens of thousands of SKUs favor learning-based approaches.

Amazon Robotics — From the Kiva Acquisition to Sparrow + Digit

Amazon acquired Kiva Systems for USD 775 million in 2012 and folded it in as **Amazon Robotics**. As of 2026, the standard stack in Amazon fulfillment centers looks like this.

- **Kiva to Hercules to Pegasus to Proteus**: The mobile drive unit lineup.

- **Proteus** (revealed 2022): Amazon's first **fully autonomous** AMR. Works alongside humans without a separated pen.

- **Sparrow**: Piece-picking arm. Vacuum + camera + ML recognition. Revealed November 2022.

- **Cardinal**: Autonomous sorter. Classifies packages into GoCarts.

- **Sequoia**: Container storage + retrieval automation. Revealed October 2023.

- **Digit** (Agility Robotics): Pilot announced October 2023. Tested the bipedal humanoid for tote moves inside Amazon facilities.

In 2024, Amazon also internalized **next-gen foundation-model piece picking** through the Covariant reverse acqui-hire. Sparrow's future is, in practice, shaped by ex-Covariant engineers.

Cube Storage — The Densest Storage Method

Stacking small totes in a 3D grid and lifting them with top-running robots delivers the highest density per square meter. The market leader is overwhelmingly AutoStore, but competitors exist.

- **AutoStore**: Norway-born. Over 1,500 sites globally. Evolved through R5, Black Line, Carousel Port.

- **OPEX iBOT**: U.S.-born. Perfect Pick + Sure Sort + Infinity G2P. The differentiator is that shuttles drive horizontally, vertically, and through the grid autonomously.

- **Element Logic**: Norway-born. Started as an AutoStore integrator, expanded with eOperator (software) / eOperator Vision.

Cube storage carries **heavy upfront CAPEX, but dominates throughput and density per square meter**. It is especially strong in urban micro-fulfillment (MFC) environments.

Sortation — The Arteries of Fulfillment

Sorters route diverse SKUs/orders into chutes or conveyor branches. Belt sorters and tilt-tray sorters were the historic standard, but modular AMR-based sorters are growing in 2026.

- **Tompkins Robotics — tSort**: U.S.-born. One AMR sorts one package into a chute. Modular setup is the differentiator.

- **Body of Knowledge AGVs**: Modular AMR sorters + WCS (Warehouse Control System) combined.

- **Numina Group**: WCS specialist. Differentiates on the orchestration layer rather than the sorter itself.

- **Honeywell Intelligrated**: Traditional large sorters + autonomous solutions combined.

- **Bastian Solutions**: Toyota Industries subsidiary. Strong in system integration.

Large-volume single-chute routing still favors traditional sorters, but AMR sorters like tSort are quickly expanding share for many-SKU, small-batch fulfillment.

Humanoids & Bipedal — Pilots, But Actually Running

Through 2024, humanoids were "YouTube demo material." From late 2025, **real warehouse/factory pilots run with measurable KPIs**. As of May 2026, the key players and pilots are as follows.

- **Agility Robotics — Digit**: Oregon, U.S. DoorDash last-mile, GXO warehouses, the Spanx pilot inside Amazon facilities. RoboFab (the first humanoid mass-production plant in the U.S.) went live in September 2023. Closer to Manufacturing-as-a-Service.

- **Figure 02**: California, U.S. Formal commercial partnership with BMW Spartanburg. Announced August 2024.

- **1X — Neo + NEO Gamma**: Norway-born. Famous for home demos. Evolved into NEO Gamma in 2025.

- **Apptronik — Apollo**: Texas, U.S. Pilot agreement with Mercedes-Benz, March 2024. Lineage from NASA Valkyrie.

- **Boston Dynamics — Atlas**: Massachusetts, U.S. Hyundai subsidiary. Revealed the **all-electric new Atlas** in April 2024, retiring the hydraulic legacy.

- **Tesla — Optimus**: Texas, U.S. In-house factory testing from 2025. External pilots remain limited as of 2026.

- **Sanctuary AI — Phoenix**: Vancouver, Canada. Teleoperation + autonomous modes. Partnership with Magna.

Humanoid ROI hinges on **whether a single form factor can handle diverse tasks**. For a single task, deterministic industrial robots remain faster and cheaper.

NVIDIA Isaac — The Standard Stack for Learning + Simulation

The biggest change in cost curves for humanoid and AMR training data came from the NVIDIA Isaac stack. As of May 2026, the core composition looks like this.

- **Isaac Sim**: Full-physical simulator on Omniverse. RTX ray tracing + accurate physics.

- **Isaac Lab**: GPU-accelerated reinforcement learning (RL) framework. Tens of thousands of parallel environments.

- **Cosmos**: World Foundation Model. Synthetic video data generator. Announced January 2025.

- **Isaac Manipulator**: A collection of pre-trained policies for 6/7-DOF manipulators.

- **Isaac Perceptor**: Multi-camera BEV perception stack.

- **GR00T N1**: Humanoid foundation model. Released March 2025.

The stack's core value is **sim-to-real**. Running on a real robot once is expensive; running 10,000 parallel environments on GPUs is feasible. Agility, Figure, and 1X publicly acknowledge using Isaac Lab for RL training.

WMS/WES/WCS — The Software Backbone Split

Fulfillment does not run just because robots are installed. **WMS (Warehouse Management System)**, **WES (Warehouse Execution System)**, and **WCS (Warehouse Control System)** — three layers — own orchestration.

- **Manhattan Associates Active Warehouse Management**: World #1 WMS. Cloud-native rewrite since 2023 made it the de facto standard.

- **Blue Yonder WMS**: Panasonic subsidiary. Luminate Logistics platform.

- **SAP EWM (Extended Warehouse Management)**: ERP integration is the strength. Strong with enterprises and manufacturing.

- **Korber Supply Chain (formerly HighJump)**: K.Motion WMS. Strong in mid-market.

- **Oracle WMS Cloud**: Cloud-native, deep integration with OCI.

- **Infor WMS**: Cloud + AI-based slotting differentiation.

WCS/WES layers have specialized vendors like **Softeon, Tecsys, Pyramid AI**. In 2026, conflicts and collaborations with robot vendors' own SW (e.g., Symbotic's SymBot OS, AutoStore's Pio) keep widening.

Collaborative Robots (Cobots) — Industrial Arms Without Safety Fences

A cobot is a robot arm designed to work safely in shared space with humans. Direct fulfillment use is less common than kitting/packing workstations.

- **Universal Robots**: Denmark-born, Teradyne subsidiary. UR3e/UR5e/UR10e/UR16e/UR20/UR30 lineup.

- **Doosan Robotics**: Korea-born. M/H/A series. Strengthened U.S. operations in 2024.

- **Techman Robot**: Taiwan-born, Quanta subsidiary. Built-in vision is the differentiator.

- **AUBO Robotics**: China-born. Global OEM supplier.

- **FANUC CRX**: Japan-born. The industrial giant's cobot line.

- **ABB GoFa + YuMi**: Switzerland-born. ABB GoFa is single-arm, YuMi is dual-arm.

ROI is decided by **workstation-level cycle time**. A line where humans kit in 10 seconds and a cobot takes 12 seconds is not worth swapping.

Korean Warehouse Robotics — Doosan, HD Hyundai, LG, CJ Logistics, Coupang

Korea's adoption picked up materially after 2024. The key players follow.

- **Doosan Robotics**: Top-tier global cobot. KOSPI IPO in October 2023. M/H/A series + Dart-Suite for AI training.

- **HD Hyundai Robotics**: Industrial robots + AMRs. Strong supply to the Hyundai group + expanding global OEM supply.

- **LG Electronics CLOi**: Service and logistics both. CLOi CarryBot, CLOi ServeBot.

- **CJ Logistics TES**: TES Innovation Center. AI-driven fulfillment operations + in-house AMRs.

- **Coupang Automated Fulfillment**: Coupang Camp. In-house sorters + AMRs + conveyor integrated operation.

- **Hyundai Glovis**: Global logistics automation solutions + in-house WMS.

- **LIGS, GIANTSTEP, Robotis**: Mid-tier in-house AMR/robotics solutions.

Korea's particularity: **a dominant share of new-build fulfillment centers**. Automation is presumed at design time, so ROI plays out faster than U.S.-style brownfield retrofits.

Japanese Warehouse Robotics — Mujin, ZMP, Daifuku, Murata, GROUND

Japan is the heartland of global material handling. The three giants Daifuku, Murata, and Toyota Industries take roughly half the market.

- **Mujin**: Tokyo-born. Deterministic motion planner MujinController. Backed by Sequoia and Toyota AI Ventures. Model-wide compatibility with ABB, KUKA, FANUC, and UR robots.

- **ZMP — CarriRo + Robot Warehouse**: Tokyo-born. CarriRo is an unmanned transporter. ZMP Robot Warehouse is a G2P solution.

- **Hitachi Industrial Products**: AMRs + industrial robots + system integration.

- **FANUC**: Yamanashi, Japan. World #1 in industrial robots. Expanding the CRX cobot line.

- **Daifuku**: Osaka, Japan. The world's largest material handling company. Covers AS/RS, sorters, AGVs.

- **Murata Machinery**: Kyoto, Japan. AS/RS + shuttle systems.

- **GROUND Inc.**: Tokyo-born. IntelliGen WES, an AI-powered warehouse operations platform.

- **Rapyuta Robotics**: Tokyo-born. ETH-derived founders. Cloud robotics.

- **Toyota Industries**: Toyota subsidiary. AS/RS + forklift automation worldwide.

Japan's strength is **dense urban fulfillment**. Deep experience designing shuttles and AS/RS into narrow, multi-floor sites.

Drone Delivery + Last-Mile — Automation Beyond the Warehouse

The warehouse is not the end. The last mile (to the customer doorstep) is automating rapidly.

- **Zipline**: U.S.-born. The global leader in medical drone delivery. Started in Rwanda and Ghana, expanded to U.S. urban P2 systems.

- **Wing (Alphabet)**: Commercial operations in Australia, the U.S., and Finland.

- **Amazon Prime Air**: Limited operations in Texas and California. The new MK30 drone arrived in 2025.

- **MissionGO**: U.S.-born. Specialized in medical and organ transport.

- **Manna**: Dublin, Ireland. European urban last-mile.

- **PabloAir**: Korea-born. Drone swarm flight + last-mile solutions.

- **Matternet**: U.S.-born. Medical P2P.

- **Skyports**: U.K.-born. Infrastructure (vertiports) + operations.

- **Starship Technologies**: Estonia/U.K.-born. Sidewalk-robot last-mile.

- **Nuro**: U.S.-born. Autonomous last-mile road vehicles.

The biggest variable in last-mile automation is **regulation (FAA/EASA/MOLIT + Aviation Safety Act) + social acceptance**. Areas with clear value, like medical, deploy first.

Grippers & Vision — The Variables That Decide Picking Success Rates

Picking-robot success rates are 80% decided by **gripper choice and vision training data**.

- **Soft Robotics (Onto Innovation)**: Soft-finger grippers.

- **Schmalz**: Germany-born. Vacuum gripper standard.

- **Schunk**: Germany-born. Mechanical finger standard.

- **OnRobot**: Denmark-born, Teradyne subsidiary. Cobot grippers + integrated vision.

- **Robotiq**: Canada-born. Cobot grippers + force-torque sensors.

- **Photoneo**: Slovakia-born. 3D vision + bin picking.

- **Zivid**: Norway-born. 3D color cameras.

- **Roboception**: Germany-born. Embedded 3D vision.

Picking success is the product of **gripper type x vision recognition accuracy x training data volume**. If any one is weak, the whole stack collapses.

Simulation + Digital Twin — What Decides Data Cost

The biggest cost in robot learning is **collecting real-world data**. A box that breaks on a single drop, or an AMR that crashes once, costs more than GPU hours. The simulation/digital twin standards follow.

- **NVIDIA Isaac Sim + Lab**: (covered above)

- **MuJoCo**: After Google DeepMind's acquisition, open-sourced. Standard for reinforcement learning.

- **PyBullet**: Lightweight open-source simulator.

- **Gazebo (Open Robotics)**: The ROS-standard simulator.

- **CoppeliaSim**: Commercial + education.

- **Drake (Toyota Research Institute)**: Accurate dynamics + MPC-friendly.

In 2026, **MuJoCo and Isaac Sim form the dual standard for training**, while Gazebo solidifies as the ROS integration default.

Real Deployment Cases — Walmart, GXO, DHL, Coupang

Finally, real fulfillment operating cases.

- **Walmart + Symbotic**: Contract to deploy Symbotic systems across all U.S. RDCs (Regional Distribution Centers) by 2025. A multi-site, multi-billion-dollar deal.

- **GXO Logistics**: World #1 3PL. Simultaneous adoption of Locus, 6 River, Boston Dynamics, and Agility — multiple robots.

- **DHL Supply Chain**: Locus + Boston Dynamics Stretch for box unloading.

- **Ocado**: U.K. grocery. In-house OSP (Ocado Smart Platform) + in-house G2P cube system. Acquired 6 River Systems in July 2023.

- **Coupang Automated Fulfillment**: In-house AMRs + some Locus deployment.

- **Amazon**: In-house Robotics stack. Estimated more than 750,000 robots in operation by 2026.

- **Alibaba Cainiao**: China-born. Geek+ + Quicktron + in-house AMRs.

The trend in deployment is shifting from **single-vendor lock-in toward multi-vendor orchestration**. The WES layer becomes increasingly important.

Adoption Roadmap — From Zero to Automated Fulfillment

Finally, a roadmap assuming you start from zero and deploy automated fulfillment.

1. **0 to 3 months**: AS-IS analysis. SKU turnover, average order lines, labor cost. Verify WMS deployment.

2. **3 to 6 months**: Choose one most urgent bottleneck. AMR pilot if picking; G2P pilot if storage density.

3. **6 to 12 months**: Pilot on a single cell. KPIs are units-per-hour (UPH) + labor savings.

4. **12 to 24 months**: Scale to multiple cells. Build WES + simulation environment in parallel.

5. **24 months and beyond**: Humanoid pilot. Start with a single task (truck unloading, tote moving).

The biggest trap is **"Let's deploy humanoids first."** Even in 2026, single-task ROI still goes to deterministic industrial robots.

Closing — In May 2026, "Warehouse Automation Is a Mosaic"

The warehouse robotics market is no longer the exclusive domain of "operators the size of Amazon." AMR/AGV, G2P, piece picking, humanoids, NVIDIA Isaac, WMS/WES, and Korean/Japanese local players all sit in the same meeting room as routine.

The key takeaways.

- AMR is effectively the standard; AGV is disappearing from new adoptions.

- G2P categorizes as AutoStore (totes), Symbotic (cases), Exotec (in between).

- Piece picking enters the foundation-model era after Covariant joined Amazon in 2024.

- Humanoids are pilots, but run with **measurable KPIs**.

- NVIDIA Isaac Sim/Lab/Cosmos overturned the training-data cost curve.

- Korea and Japan are strong in new builds and dense urban environments.

Most important: **"No single vendor fits every seat."** Honestly measure your warehouse's SKU mix, turnover, labor cost, and whether it's new-build — then design the layered stack on top.

References

- Symbotic official site: https://www.symbotic.com/

- AutoStore official site: https://www.autostoresystem.com/

- Locus Robotics official site: https://locusrobotics.com/

- Berkshire Grey official site: https://www.berkshiregrey.com/

- GreyOrange official site: https://www.greyorange.com/

- Geek+ official site: https://www.geekplus.com/

- Hai Robotics official site: https://www.hairobotics.com/

- Exotec official site: https://www.exotec.com/

- Amazon Robotics official site: https://www.aboutamazon.com/news/operations/amazon-introduces-new-robotics-solutions

- Agility Robotics — Digit: https://agilityrobotics.com/products/digit

- Figure 02 official site: https://www.figure.ai/

- 1X Technologies — Neo: https://www.1x.tech/

- Apptronik — Apollo: https://apptronik.com/

- Boston Dynamics — Atlas: https://bostondynamics.com/atlas/

- Tesla Optimus: https://www.tesla.com/AI

- Sanctuary AI official site: https://www.sanctuary.ai/

- NVIDIA Isaac Sim: https://developer.nvidia.com/isaac/sim

- NVIDIA Isaac Lab: https://developer.nvidia.com/isaac/lab

- NVIDIA Cosmos World Foundation Model: https://www.nvidia.com/en-us/ai/cosmos/

- Manhattan Associates: https://www.manh.com/

- Blue Yonder: https://blueyonder.com/

- SAP EWM: https://www.sap.com/products/scm/extended-warehouse-management.html

- Universal Robots: https://www.universal-robots.com/

- Doosan Robotics: https://www.doosanrobotics.com/

- HD Hyundai Robotics: https://www.hd-hyundairobotics.com/

- LG Electronics CLOi: https://www.lge.co.kr/business/robot/cloi

- CJ Logistics TES: https://www.cjlogistics.com/en/tes/main

- Mujin official site: https://www.mujin.co.jp/

- ZMP CarriRo: https://www.zmp.co.jp/products/logistics-support-robot-carriro

- Daifuku official site: https://www.daifuku.com/

- Murata Machinery: https://www.muratec.net/

- GROUND Inc.: https://groundinc.co.jp/

- Zipline official site: https://www.flyzipline.com/

- Wing (Alphabet): https://wing.com/

- Amazon Prime Air: https://www.aboutamazon.com/news/transportation/prime-air-delivery-drone-mk30

- Covariant (Amazon acquisition announcement): https://www.aboutamazon.com/news/operations/amazon-covariant-ai-robotics