LabHub

Blog

AI Warehouse Robotics & Fulfillment 2026 Deep Dive - Symbotic, Berkshire Grey (SoftBank), Locus Robotics, AutoStore, GreyOrange, Geek+, Hai Robotics, Amazon Robotics, NVIDIA Isaac, Doosan Robotics, ZMP Complete Guide

2026-05-16 · 16 min read · #warehouse-robotics#fulfillment-ai#symbotic#berkshire-grey#locus-robotics#autostore#greyorange#geek-plus#hai-robotics#amazon-robotics#nvidia-isaac#doosan-robotics#english

한국어English日本語

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.

The representative AMR/AGV players are as follows.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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.

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

Comments

No comments yet.

Sign in to leave a comment