Retail and consumer goods
How retailers and consumer goods brands use NVIDIA AI for store vision, smart carts, shopping and catalog agents, product digital twins and delivery routing, and where shopper privacy, content accuracy and simpler off-the-shelf tools should shape the plan.
The problem
Most grocery spending still happens in physical stores, yet retailers often find out about empty shelves, wrong planograms and shrink only after the fact. Online, product pages are frequently thin: suppliers send short titles, a few attributes and one photo, which weakens search and recommendations in every market and language.
Consumer goods brands have a content problem of their own. One product line sold in several sizes, variants and languages can need thousands of approved images, and every reshoot costs time and money. Behind the store, fulfillment teams plan picking, replenishment and last-mile routes against tight delivery windows.
All of this runs on thin margins and sharp seasonal peaks, so a system has to pay for itself quickly and stay dependable on the busiest trading days.12
The approach
In stores, NVIDIA positions Metropolis as the base for camera analytics such as asset protection, inventory monitoring and quality inspection. Part of the work runs on edge devices: Instacart's Caper Cart combines five cameras, a certified scale and a Jetson Orin NX module to recognize items as shoppers place them in the basket, while Instacart's ad and item ranking models run on Hopper GPUs in the cloud with NVIDIA Dynamo.
For digital commerce, NVIDIA publishes retail blueprints on GitHub, including a multi-agent shopping assistant that uses retrieval over the product catalog and NIM microservices, and a catalog enrichment workflow that uses Nemotron models to write localized titles, descriptions and FAQs from product images. For brand content, Unilever keeps one approved 3D twin per product in Omniverse with OpenUSD, an open scene standard created at Pixar. For planning, NVIDIA offers cuOpt for routing, scheduling and fulfillment problems.
A simpler path is often enough. Analytics built into the store's camera or point-of-sale system can cover queue counts and basic shelf alerts, a hosted e-commerce search service handles discovery for a mid-sized catalog, and open source CPU solvers handle routing for a small fleet. GPU systems tend to pay off with many stores, very large catalogs or high request volumes.123456
Conceptual architecture
Diagram as a list
Applications & solutions
- Store cameras, smart carts and shelf sensorsCapture what happens at the shelf, in the basket and at checkoutConnects to In-store edge inference (Jetson, Metropolis pipelines)
- Shopping assistant and catalog enrichment agents (Blueprints, Nemotron, NIM)Answer shopper questions and draft product content for human approvalConnects to Web, app, in-store screens and marketing channels
- Product and store digital twins (Omniverse, OpenUSD)Hold approved 3D products and store layouts for imagery and planningConnects to Web, app, in-store screens and marketing channels
- Fulfillment and route optimization (cuOpt)Plans picking, replenishment and delivery routes
- Web, app, in-store screens and marketing channelsDeliver recommendations, assistant answers and approved imagery to shoppers
Inference & runtime software
- In-store edge inference (Jetson, Metropolis pipelines)Recognizes products and shelf states close to the cameras and sends events, not raw videoConnects to Unified commerce data (orders, inventory, store events)
- Search, ranking and ad models on GPUs (Dynamo)Scores products and sponsored items for each page, cart or screenConnects to Web, app, in-store screens and marketing channels
Operations & orchestration
- Unified commerce data (orders, inventory, store events)Joins online and in-store signals under shared product and store identifiersConnects to Search, ranking and ad models on GPUs (Dynamo), Shopping assistant and catalog enrichment agents (Blueprints, Nemotron, NIM), Fulfillment and route optimization (cuOpt)
Technologies and their roles
NVIDIA Metropolis3
Store vision analytics
NVIDIA names asset protection, inventory monitoring and quality inspection among Metropolis uses in retail and consumer goods.
NVIDIA Jetson1
On-cart and in-store edge computing
Each Instacart Caper Cart uses a Jetson Orin NX module for real-time sensor fusion.
NVIDIA Dynamo1
Ranking and recommendation serving
NVIDIA reports that Instacart moved ad ranking workloads from CPUs to Hopper GPUs with Dynamo.
NVIDIA Blueprints3
Starting code for commerce agents
NVIDIA publishes retail shopping assistant, catalog enrichment and agentic commerce reference code on GitHub.
NVIDIA Omniverse3
Product and store digital twins
Unilever builds product twins on Omniverse, and NVIDIA lists store layout and warehouse simulation as retail uses.
NVIDIA cuOpt3
Fulfillment and delivery optimization
NVIDIA names routing, scheduling, warehouse planning and last-mile delivery as cuOpt uses in retail.
What you need first
- A clean product master with stable IDs, attributes, images and the rights to use them
- A camera and sensor inventory per store, plus network capacity for edge devices
- A documented purpose, privacy notice, signage and legal basis for in-store video, agreed with legal teams and works councils where required
- Labeled examples of shelf states, products and checkout events from your own stores
- Order, inventory and delivery data with identifiers that match across online and store channels
- 3D product data (CAD or scans) and a partner able to build accurate twins, if product imagery is in scope
- Operations skills to update models on many stores and devices without disrupting trading
Risks and how to reduce them
- Shopper privacy and biometric rules for in-store cameras
- Analyze products and shelves rather than people where possible, avoid face recognition without a clear legal basis, blur or discard personal images, keep retention short and post clear notices.
- Generated product content that is wrong or breaks labeling and advertising rules
- Keep human approval for generated titles, claims and images, and check them against brand policy and product labeling law before publishing.
- Shopping agents that buy or promise things the customer did not intend
- Require explicit confirmation before checkout, limit what an agent can change, and log every agent action.
- Unfair loss prevention alerts
- Treat an alert as a prompt for a staff check rather than an accusation, and audit alert rates across stores and customer groups.
- License terms of reference code and the models it calls4
- Confirm each component's license before production; the shopping assistant code is Apache-2.0, but the models and services it calls carry their own terms.
Documented examples
Unilever · Consumer goods (marketing content production)
Unilever: product digital twins for marketing imagery with Omniverse and OpenUSD
Unilever builds photoreal 3D twins of its products with NVIDIA Omniverse and OpenUSD, working with creative technology partner Collective World, and renders marketing images from them instead of running repeated photo shoots. Unilever and NVIDIA report imagery made twice as fast at half the cost.
In production
Instacart (Maplebear Inc.) · Grocery retail technology
Instacart: Caper smart carts on Jetson and GPU ranking with Dynamo
Instacart runs item recognition on its Caper smart carts with NVIDIA Jetson Orin NX modules and moved online ad and item ranking to NVIDIA GPUs with Dynamo. Published results include 65 percent lower whole-page ranking latency and an incremental sales lift above 1 percent in A/B tests.
Scaling
Related
Sources
- Instacart case study (Retail/Consumer Packaged Goods) (opens in a new tab)
- Unilever customer story (Retail / Consumer Packaged Goods) (opens in a new tab)
- NVIDIA retail and CPG industry page (opens in a new tab)
- NVIDIA AI Blueprint: Retail Shopping Assistant (GitHub README) (opens in a new tab)
- Retail Catalog Enrichment Blueprint (GitHub README) (opens in a new tab)
- Alliance for OpenUSD (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.