Why Do Retail Robots Still Drop Bottles on Real Shelves?

Introduction

A robot that can dance or perform a polished demonstration is not necessarily ready to complete a retail job. In 2025, Hongqi Chain, a major convenience-store operator in southwest China, announced a project to develop a robot-based unmanned vending system. The announcement briefly excited investors, and the company’s share price rose by about 35% from before the announcement to late April that year. After around 16 months of development, however, Hongqi announced that the project would be discontinued because the system had not reached the threshold for large-scale commercial deployment.

The gap between the lab and the store

The company had completed hardware and software tests in a laboratory environment. The real store exposed a much harder problem. Condensation on cold glass bottles can make grasping unreliable. Different packaging materials require different force, angles, and handling strategies. Narrow aisles, changing light, nearby customers, and store employees also complicate perception, collision avoidance, and motion planning.

The challenge is therefore not simply making a robotic arm pick up one item. A commercial system must recognize the product, understand the order, plan a route, grasp the item, verify the result, and deliver it consistently. A failure at any point can interrupt a transaction or damage the product. Hongqi ultimately shifted its focus toward scaling a 24-hour cloud-supervision service, a sign that remote monitoring and human fallback remain more practical than fully autonomous operation.

What retail robots actually do today

• Retail pods are controlled environments. They usually carry a limited selection of drinks, coffee, and other standardized products. Staff must arrange the shelves in advance, because a robot may not recover from a misplaced or missing item.
• The robot often performs only retrieval. After a customer places an order, it moves to the shelf and brings back the product. Restocking, charging, and fault handling still require on-site workers.
• Engagement can be more valuable than productivity. Robots greet visitors, describe products, and create a reason to take videos. This can generate attention in malls and tourist sites, but it does not mean the robot has replaced a clerk.
• The economics remain unproven. A complete retail pod may approach the cost of a high-end industrial system, while staffing, maintenance, charging, and site expenses continue. A small product range and weak repeat purchasing make payback difficult.

Why companies keep deploying them

Retail is an attractive intermediate scenario. It is more open than a factory but more tolerant of mistakes, and a robot placed in a busy landmark can function as both a test platform and an advertising display. Tourism operators may value the attention and social-media exposure, while robot makers gain real-world data.

Yet visibility is not the same as operating value. Commercialization requires safe operation around people, faster task completion, broader generalization, and a total cost below that of human labor. Industry participants cited in the source estimate that replacing retail employees may still take two to five years. The likely route is gradual: controlled vending pods first, then counter service, semi-open stores, and eventually conventional supermarkets. Hongqi’s experience is a useful reminder that embodied intelligence is not proven by one successful demo, but by performing hundreds of messy real-world tasks every day at a sustainable cost.

Source: QbitAI