There are two ways to develop retail robotics. Build in a lab until the system is perfect, then meet the public. Or meet the public early, in demanding venues, and let reality set the engineering agenda.

We chose the second path. Robox robotics went live at Paris Saint-Germain’s Parc des Princes and at the Museum of the Future in Dubai, and have now processed more than 30,000 transactions with real customers in real conditions. Those venues were not soft launches. A stadium tour and a flagship museum are among the harshest environments retail technology can face: massive footfall surges, zero staff tolerance for machine problems, customers of every age and language, and a reputational spotlight where any failure is public.

Here is what that education taught us, lessons now built into every unit we deploy.

Lesson one: throughput is a cliff, not a curve

In a normal store, demand arrives as a flow. In a stadium or museum, it arrives as a wall: a tour group releases, a match ends, and a quiet unit faces a hundred people in minutes.

The engineering consequence is that average performance is almost irrelevant. What matters is behavior at peak, cycle time under continuous load, queue psychology at the interface, payment reliability when transactions stack. We redesigned interaction flows so that the slowest step in the cycle is never the customer standing confused. If a first-time user cannot complete a purchase without help, the design has failed, because there is no help.

That discipline, design for the surge, and the quiet hours take care of themselves, now shapes every product decision.

Lesson two: uptime is a brand promise, not a metric

When a robot represents Paris Saint-Germain or the Museum of the Future inside their own venue, a frozen screen is not a support ticket. It is the venue’s brand failing in front of its visitors.

That standard forced our operations model into its current shape: remote-first, preventive, and continuous. GateX telemetry watches every unit’s mechanical health, error patterns, and transaction success rates, and flags anomalies before they become visible failures. Maintenance is scheduled from data, not from breakdowns. The philosophy is simple to state and demanding to build: the venue should learn about a problem from us, after it has been handled, never from a visitor.

Zero churn across our host locations is the commercial evidence that the model works. Venues renew because the machines show up every day.

Lesson three: the robot is the marketing

We expected the robots to transact. We underestimated how much they would attract. In both venues, the units became moments in the visit, people stop, watch, film, and share. A robot working in public is a spectacle in a way a shelf can never be, and that spectacle drives commercial results directly: footfall converts because the machine itself creates the stop.

The measurable version of this surprised us most. At Parc des Princes, thousands of visitors have voluntarily engaged with the experience deeply enough to leave contact details, engagement rates that conventional retail activations rarely approach. For venue partners, this reframed what the unit is: not just a point of sale, but an attraction that happens to sell, and a data channel that happens to delight.

This lesson carries straight into the autonomous store. Machinery that operates gracefully in public is not just infrastructure; it is the storefront’s best advertisement.

Lesson four: every venue is a different store

The PSG deployment and the Museum of the Future deployment run on the same platform and behave like different businesses. Different visitor rhythms, different peak hours, different product velocities, different languages at the screen. A stadium sells to emotion on a schedule set by fixtures; a museum sells to wonder on a schedule set by tourism seasons.

The operational conclusion: configuration must be a platform capability, not a project. GateX now treats assortment, pricing, content, and language as per-unit variables managed remotely, which is exactly the muscle a multi-location autonomous fleet needs. Learning it at two flagship venues, with two demanding partners, was the fastest possible teacher.

Lesson five: trust is transferable

The deepest lesson is strategic. Retail robotics faces a cold-start problem: partners hesitate to host machines that haven’t proven themselves, and machines can’t prove themselves without hosts. Flagship venues broke that loop for us. When a system has run publicly at Parc des Princes and the Museum of the Future, venues that audit their partners hard, every subsequent conversation starts from demonstrated reality rather than promises.

The 30,000 transactions are not a vanity number. They are the compressed answer to the question every prospective partner rightly asks: has this worked, with real people, where failure would have been visible? It has. New locations now go live every month on the strength of that answer.

What carries forward

The autonomous store is a bigger machine than a personalization robot, but it inherits everything above: engineering for the surge, operations that prevent rather than react, machinery designed to be watched, configuration as a native capability, and proof built in public.

Labs teach you what a robot can do. Thirty thousand strangers teach you what a robot must do. We recommend the strangers.