Robotics for delivery-first grocery
Robotic picking and shelf intelligence for small, delivery-first grocery stores. Starting in Germany.
Demonstrated in simulation.
Building toward physical integration, then a pilot in Germany.
DRAY in action. Recorded demonstrations from our simulated store.
Intelligence, in action
See how a shelf observation becomes a clear task, a robotic action and a checked result.
Wrist camera 
Selected position
Follow the product from the supply tray to its selected shelf position.
Keep the shelf ready for the next order.
Recorded in our simulated store. Select a stage to explore the observation, action and result.
36 / 36Restock trials on unseen product textures
18 held-out products · 2 shelves each
18 / 18Restock & relocate trials
Known shelves
Simulation evaluations · September 2026. Physical integration comes next.
A closer look
Shared aisles. Changing shifts. Orders collected and shelves restored. Four short simulation stories.
Simulation scenarios exploring the DRAY vision.
Explore the capabilities behind the work in the interactive demo.
Results so far · simulation
These clips show the learned policy at 1× speed, as recorded. The table reports evaluations of the current checkpoint; the clips show an earlier checkpoint.
| What was measured | Result |
|---|---|
| Restock on shelves with unseen product textures18 held-out real products, 2 shelves each | 36 / 36 |
| Restock and relocate on known shelves | 18 / 18 |
| Three-stop replenishment trip | 3 / 3 stops |
| Navigation among simulated shoppersROS 2 / Nav2 navigation stack, not the learned policy | 20 / 20 arrivals · 0 collisions |
SmolVLA-based policy trained on 38,000+ simulated episodes. Table: evaluations of the current checkpoint (Sep 2026); clips: recorded runs of an earlier checkpoint of the same policy. Simulation results; physical integration is the next step.
The proof · Recorded in simulation
Follow one recorded replenishment task from a shelf image to a verified result. The realogram connects product identities and positions to the robot’s next action.
See → Decide → Act → VerifyDRAY reads product identities, positions and visible facings from the shelf image. The realogram makes that information usable for shelf audits, task selection and navigation.
Shelf recognition · recorded simulation
Compare what is on the shelf with what belongs there. Here, DRAY finds two facings where three are expected and selects the position to replenish.
Shelf insight becomes a restocking taskA precise target connects shelf understanding to the robot’s next action.
The robot takes the selected product from its tray, reaches the target and places it on the shelf. Shelf intelligence becomes a physical action.
Robotic replenishment · simulationThe selected product moves from supply tray to shelf.
A fresh shelf image checks the result. In this recorded task, the missing facing is restored and the shelf record is updated.
Before and after, connected by a task
See the shelf. Do the work. Check the result.
See AI in actionInside DRAY
Step inside the simulation. Reveal an AI view of the world, then explore the intelligence behind every action.
Products to handle. People to move around. A place to understand before taking the next step.
Interactive illustration using our simulation scene. AI view and route are visual explanations.
Trigger an illustrative shelf gap to follow the decision through the graph.
Simulation store structure and modelled product relationships. Product dots are placed schematically within each bay. Shelf-gap scenario is illustrative.
A dedicated grocery format for robotic fulfillment, with a lean team supporting the operation.
Explore the partnershipApply the same capabilities to replenishment, misplaced products and repetitive picking in conventional stores.
See the capabilitiesRead product positions and visible facings. Identify gaps, compare shelf layouts and check completed work.
Explore shelf intelligenceThe robotics platform
One mobile platform connects shelf understanding with product handling. Explore the capabilities in our simulated store.
Docked at the bottom shelf, the robot takes a box from its supply tray and sets it into the one empty facing beside its siblings. From a supply tray to the right shelf position.
The robot takes a misplaced can off the top shelf into its tote, then sets it into its empty slot one level down. Two recorded segments, joined at the tote. Put misplaced products back where they belong.
The robot takes a bottle off the shelf into its tote, then lifts it from the tote into the order bin. Two recorded segments; the drive between them is cut. The core movement behind robotic order fulfillment.
Recorded DRAY demonstrations · Simulation
For retailers
We’re seeking a retail partner for small, delivery-first stores in Germany. Shape the operating model now; a physical pilot follows integration and validation.
You own the store. DRAY provides the robot and operating software as a monthly service.
Shape the assortment, routine tasks and human support with us before a physical pilot.
Cost per order, shelf availability and human interventions. Measured together at the first site.
01 · Labour
An hour of retail work costs a German employer €31.18. The statutory minimum wage rises 17.6% between 2024 and 2027, with night and Sunday surcharges on top. About 122,000 retail positions were vacant on average in 2024. Keeping one person on duty around the clock in a small store takes about 5.3 full-time staff, roughly €185,000–195,000 a year. A small store cannot carry that.
Sources: Eurostat labour cost survey 2024 · Mindestlohnkommission 2025 · HDE/KOFA 2025 · DRAY estimate02 · Format
Żabka runs about 13,000 stores of 70–120 m²; 18 million Poles live within 500 m of one, and a regular store pays back in 11–12 months. Its checkout-free Nano company still employed 68 full-time staff, head office included, for about 50 locations, and a Nano store pays back in 3–4 years. Removing the cashier did not remove the work. Restock, pick and verify is what DRAY does.
Sources: Żabka Group results 2024–H1 2026 · Supermarktblog, May 202603 · Hours
Hessen (up to 120 m²), Bavaria and Baden-Württemberg (up to 150 m²) let automated stores of everyday goods open on Sundays without staff. At the unstaffed chain teo, Sunday brought 25–30% of weekly sales. A robot keeps the shelves ready for the hours your staff cannot cover.
Sources: HLöG Hessen 2024 · BayLadSchlG 2025 · LadÖG BW 2026 · Supermarktblog, Feb 2024Sunday rules differ by state, are new, and limit what may be restocked on Sundays; DRAY plans every site with local counsel.
04 · Delivery
REWE now picks one-hour orders inside regular supermarkets, one store per roughly 3 km, and expects one to two extra pickers per store. Flink says it is EBITDA-profitable on about 160 hubs of 3,000 products each: the assortment of a small store. Central robotic warehouses need heavy capital: Picnic's Oberhausen centre cost €150m, and Kroger closed three Ocado sites in 2026. A small store on ordinary shelving uses the building you already have.
Sources: REWE and Supermarktblog, 15 Sep 2026 · Prosus, 3 Mar 2026 · LOGISTIK HEUTE, Jun 2025 · Ocado HY2605 · Your stores
Across Europe, 8.6% of items were missing when a shopper looked for them (2002–03), and about 70% of the causes sat inside the store. Two-thirds of inventory records are wrong; at one UK grocer, a store-wide audit was followed by about 11% more sales over two months. Shelf-scanning robots find the gaps. People still fill them. DRAY closes the loop.
Sources: Gruen & Corsten 2003, ECR Europe · DeHoratius & Raman 2008 · UK grocer field study 2025With a retail partner in Hessen, Bavaria or Baden-Württemberg. Agreed KPIs: cost per completed order, shelf availability, human interventions. Assortment grows step by step.
The same robot audits, replenishes and verifies while the store is closed, and picks delivery orders during the day.
A monthly fee per robot and software. Drugstores and convenience formats next.
Build with us, from the start
Start with the workflows, assortment and operating constraints that matter to your business. Agree what a future pilot must prove, before choosing a site and committing to deployment.
Agree the measures now. Validate them in a future pilot.
For early-stage investors
Seed round: €2–3M for 18 months, to put DRAY on real hardware and into two to three stores in Germany. Simulation-proven today; the store is the proof point, the service is the business.
SimulationPhysical integrationPilot
36,565 food stores in Germany, 7,900 of them under 400 m². 4,496 dm and Rossmann drugstores, 16,601 pharmacies, 14,018 petrol stations. At €3,000 per robot per month, a bottom-up German market of about €3.7bn a year, 4.8% of what German retail spends on labour. Europe: roughly €17bn.
Sources: Fee is a working assumption set against robot subscription prices (Simbe 2020, MiR, 1X); store counts EHI 2024, dm/Rossmann 2025, ABDA 2025, bft 2025, Eurostat 2024.At €3,000 a month one robot costs the same as 0.7 of a full-time retail worker, or 96 staff hours a month. Hardware runs €25–50k per robot at today's component prices: about 49% gross margin and hardware payback in about 17 months. A minimum-wage picker will cost about €22 per hour worked in 2027.
Sources: DRAY estimates from public prices and Eurostat labour costs.Robot subscriptions grew 31% in 2024 while robot sales grew 9%. Simbe, which sells its shelf scanner as a subscription, has 3,000+ units under contract, and every one of them hands the fix to a person. Store-based delivery is back at REWE, and three German states have opened Sundays to automated stores.
Sources: IFR 2025 · Simbe, Sep 2026 · REWE, Sep 2026Beyond grocery: drugstores, convenience and petrol-station shops, pharmacies' over-the-counter shelves.
Who is building DRAY
A team that has shipped AI in production for a decade, robots included. DRAY is our next system.
CEO & Founder
Ten years of production AI where a wrong output costs money or lives. Built computer vision for the Asensus Senhance surgical robot, including eye-tracking camera control, and worked on autonomous-robot AI in the SoftServe team at MBZIRC 2020.
CTO
Hands-on engineer who has run ML platforms and agentic systems under real load; former CTO at Neurons Lab. At MBZIRC 2020 he worked on perception and tracking for autonomous drones and ground robots in the same SoftServe team.
Senior AI Engineer
Came to AI from bank risk management, the discipline of pricing what can go wrong before it does. At DRAY he builds the evaluation side: the gates, checks and simulation runs that decide when a robot task counts as done.
Deep Learning AI R&D Engineer
Trains and fine-tunes the models. At DRAY that means the vision-language-action policies behind the robot’s arm and the simulation data pipeline that feeds them.
Chief Business Development Officer
A decade across sales, customer success and go-to-market, through a B2B SaaS exit. Owns retail partnerships: the first conversation with a grocer and the pilot terms that follow.
Product Manager
Turns retailer workflows into product requirements: the assortment, the tasks to automate and the success thresholds a future pilot must prove.
A few practical questions
Have a specific store or workflow in mind? Tell us about it.
You can inspect recorded AI and robot tasks in simulation: shelf recognition, replenishment, relocation and picking. Physical integration comes next, supported by investment and an expanded team. A store pilot follows integration and validation.
A monthly service: the robot and its operating software. You own and run the store. We help with the first sites, agree the KPIs and publish the results with you.
Delivery-first grocery is our initial focus. Shelf work alongside existing store teams and shelf analytics are potential later applications of the same capabilities.
It structures the visible shelf: product identities, positions and facing counts. Compared with the expected layout, that supports gap detection, misplaced-product checks and verification after a task. It also provides product landmarks for navigation.
The pilot concept puts people in charge of maintenance, oversight and exceptions while robots handle routine shelf and order work. The first site will establish the right balance of scheduled visits, remote support and on-site intervention.
Extended service windows, including Sundays where permitted, are part of the concept. Opening, staffing and delivery arrangements depend on the location and operating model; these would be assessed with the pilot partner.
Start with your workflows, assortment and operating goals. We are seeking a co-development partner to shape requirements and success criteria now. A physical pilot comes later, after integration and validation; site, responsibilities and commercial terms would be agreed together.
Let’s take the next step
Shape the workflow and success criteria for the first DRAY store, starting with your retail context.
Explore co-developmentHelp expand the team and fund the move from simulation to physical integration.
Start an investor conversation