Logistics and warehousing
How logistics providers, retailers and shippers use AI and NVIDIA technologies in distribution centers and transport: route and task optimization, warehouse twins, robot fleets trained in simulation, camera-based label reading and planning agents, plus when existing systems are enough.
The problem
Logistics runs on thin margins and firm delivery promises. Volumes swing with seasons and promotions, customers expect narrower delivery windows, and one late truck or blocked dock door ripples through the rest of the day. Many operators still replan routes and shifts in batches, so they respond to disruption well after it starts.
Inside distribution centers, forklifts, autonomous mobile robots and conveyors share aisles with pickers. Adding automation or moving racking in a building that ships every day is risky, and labels that are torn, wrinkled or half hidden get misread, which turns into wrong shipments and returns.
The data needed to see the whole network is split across warehouse, transport, yard and carrier systems, often owned by different companies, so answering a planner's question about stock, capacity or a late order can take hours of manual lookups.
The approach
Routing, replenishment and task allocation are optimization problems. cuOpt, NVIDIA's open source solver that runs on GPUs, handles vehicle routing plus linear and quadratic models, while its mixed-integer support is still in beta. Its product page lists supply chain management, fleet management, last-mile delivery and field dispatch as uses, and names Lowe's for re-optimizing shipping routes during disruptions and Blue Yonder for last-mile delivery. NVIDIA's supply chain page adds Omniverse twins of warehouses and fulfillment centers for testing layout and throughput changes before construction, Metropolis vision AI for reading damaged or partly hidden labels, and agents such as warehouse assistants and demand planners built with NeMo, NIM and Nemotron.
Robot fleets are rehearsed in simulation first: NVIDIA's retail page names Isaac Sim with Omniverse for training autonomous robots and simulating warehouse operations. In January 2025 KION Group, Accenture and NVIDIA showed a warehouse twin in which KION's warehouse management software assigns tasks to virtual robots, so operators can size robot and worker numbers without disturbing the live site; KION's release describes a showcase and names no operating site. PepsiCo applies Siemens Digital Twin Composer, built on Omniverse libraries, to selected US warehouses as well as plants.
A lot of logistics work needs no GPU. Routing for a regional fleet of modest size is usually handled by the optimizer inside a transport management system or a CPU solver, fixed scanners and tunnel readers cope with clean labels, and event-based simulation answers most questions about slot assignment and shift sizes. GPU optimization and 3D twins pay off for national networks, frequent replanning during the day, or robot fleets that need training on the specific building.12345
Conceptual architecture
Diagram as a list
Applications & solutions
- Order, warehouse and transport management systemsHold orders, stock, vehicles, time windows and task queuesConnects to Route and task optimization (cuOpt), Warehouse twin (Omniverse libraries), Planning assistant agents (NIM, Nemotron)
- Warehouse twin (Omniverse libraries)Tests layouts, throughput and automation before changes on siteConnects to Robot fleet simulation and training (Isaac Sim)
- Dispatch, yard and floor executionCarries out plans and reports actual times and exceptions backConnects to Order, warehouse and transport management systems
Models & frameworks
- Robot fleet simulation and training (Isaac Sim)Trains and checks mobile robot behavior against the virtual buildingConnects to Dispatch, yard and floor execution
Inference & runtime software
- Dock and aisle cameras with vision AI (Metropolis)Reads labels and tracks movement across camerasConnects to Dispatch, yard and floor execution
- Planning assistant agents (NIM, Nemotron)Answer planner questions and frame problems for the optimizerConnects to Route and task optimization (cuOpt)
Accelerated computing
- Route and task optimization (cuOpt)Builds and rebuilds delivery routes and picking or replenishment assignmentsConnects to Dispatch, yard and floor execution
Technologies and their roles
NVIDIA cuOpt1
Routing and task allocation
Open source GPU solver whose product page lists fleet management, last-mile delivery and field dispatch, with Lowe's and Blue Yonder named as users.
NVIDIA Omniverse2
Warehouse and fulfillment center twins
NVIDIA's supply chain page names Omniverse for testing warehouse layout and throughput changes, and PepsiCo's warehouse twins run on Siemens software built on its libraries.
NVIDIA Isaac3
Robot fleet simulation
NVIDIA names Isaac Sim with Omniverse for training autonomous robots and simulating warehouse operations.
NVIDIA Metropolis2
Label reading and floor visibility
Vision AI for reading damaged or partly hidden labels and for multi-camera tracking in warehouses.
NVIDIA NIM2
Serving models for planning agents
NVIDIA's supply chain page names NIM, with NeMo and Nemotron, for warehouse assistant and demand planning agents deployed through AI Enterprise.
What you need first
- Clean order, address and time-window data, plus travel times or road distances from a mapping provider
- Business rules for priorities, driver hours, vehicle limits and dock capacity written down so they can become solver constraints
- Interfaces to the warehouse and transport management systems that will carry out the plans
- Building drawings, rack and conveyor layouts and robot specifications if a warehouse twin is planned
- Operations research skills to model problems and judge the quality of solutions
- Data-sharing terms with carriers, third-party logistics providers and customers whose data enters the models
Risks and how to reduce them
- Tracking of pickers and drivers
- Measure flows and zones rather than individuals where possible, keep footage and telematics only as long as the purpose needs, and agree the rules with staff and their representatives.
- Plans that break real-world rules
- Encode driver hours, dock limits and load rules as hard constraints, let dispatchers approve plans, and compare planned with actual times every week.
- Beta solver features in production1
- cuOpt's mixed-integer support is in beta; cross-check results against an established solver before relying on them.
- Robots, forklifts and people in shared aisles
- Keep certified safety scanners, speed zones and separated traffic routes; a robot trained in simulation still needs on-site safety acceptance.
- Exposure of partner shipment data
- Shipment and customer data from carriers and clients often falls under contract limits; minimize what enters shared models and control access per partner.
Documented examples
PepsiCo, with Siemens · Food and beverage manufacturing and logistics
PepsiCo and Siemens: plant and warehouse twins with Digital Twin Composer on Omniverse
PepsiCo is converting selected US plants and warehouses into 3D digital twins with Siemens Digital Twin Composer, which Siemens builds on NVIDIA Omniverse libraries. PepsiCo and Siemens report a 20 percent throughput gain on the first deployment; the program is in early pilots.
Pilot
Related
Sources
- NVIDIA cuOpt (product page) (opens in a new tab)
- NVIDIA intelligent supply chain (industry page) (opens in a new tab)
- NVIDIA retail and CPG industry page (opens in a new tab)
- KION teams with NVIDIA and Accenture to optimize supply chains with AI-powered robots and digital twins (opens in a new tab)
- PepsiCo announces industry-first AI and digital twin collaboration with Siemens and NVIDIA (opens in a new tab)
- NVIDIA Metropolis (product page) (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.