Two Companies Partner on Embodied AI Robots for Supermarket Inventory Checks

The robot hardware moves along aisles, avoiding shoppers and carts, while reading electronic shelf labels and checking whether items are in the right place. During restocking, a robotic arm reaches into shelves only 30 to 40 centimeters high to pick up soft bags and place them neatly. Recognition accuracy, operational safety and equipment cost are all factors that directly affect a retailer's bottom line.

Hanshow, a retail digitalization provider, serves more than 550 retail customers in over 80 countries and has installed electronic shelf labels in roughly 70,000 stores. These labels are connected to real-time product and price data, forming a physical coordinate system that a robot can read without building a fresh map of every store.

X-Era Lab, incubated by Sun Yat-sen University's HCP lab and Peng Cheng National Laboratory, contributed its native world action model developed on a 4D spatiotemporal architecture. The model is designed to handle occlusion, changing light and product state variations, allowing the robot to predict changes and execute actions. With around 1 billion parameters, the model is said by the report to outperform much larger mainstream systems on more than ten benchmarks.

The partnership was formed because retail stores are viewed as an underappreciated test field for embodied AI, sitting between highly structured factories and unstructured homes. A major bottleneck in retail has been inventory frequency: full-store counts are usually done only once or twice a year, since they have to be performed after closing hours to avoid interference. The report cites industry data showing that mismatches between recorded and actual stock are common, and that increasing count frequency could significantly improve real-time inventory control and loss prevention.

Hanshow's electronic labels give the model a starting assumption of what should be on each shelf, so the robot only needs to confirm against a closed set of possibilities, which reduces edge-computing burden and latency. The labels also provide native semantic coordinates, removing the need for manual annotation after SLAM mapping. In regions with strict data regulations such as Europe, relying on label priors and on-device small models enables local data processing, turning compliance into a technical advantage, the report said.

The partners plan to use real store operations as a continuous training loop. Each inspection, count or handling error becomes a feedback signal to improve the model. Hanshow has also built a digital-twin layer that maps products, shelves, equipment, inventory and store status into a unified digital space, giving robots access to constantly updated prior knowledge of the environment.

Hanshow's inspection robots have already entered proof-of-concept tests at shopping malls in China and overseas, according to the report. The companies say they want to move beyond one-off demos and achieve a commercially viable production system, rather than simply placing a general-purpose model in a store.

Chen Tianshui, CTO of X-Era Lab, said in the report that the retail track is difficult to sustain with conceptual stories alone because physical scenes are highly fragmented; the single task of inventory counting can be broken into many evaluation levels, and technical capability is easily falsified in actual business. Tong Liang, general manager of Hanshow's robot product line, said a retail venue is a typical semi-structured scene: shelf layout and basic operating rules are fixed, but variables such as product changes and customer movement make it a suitable data source for training embodied robots, while reducing the cost of trial and error.