Editor’s note: This article was written by Lucia, a TechNode reporter.
On Sept. 23, SenseTime launched SenseMart OS at its Shanghai headquarters, positioning it as a physical operating system for retail. The system connects heterogeneous robots, retail equipment, and commercial operations, offering operators, brands, and venues a deployable, operable, and replicable embodied-retail solution. At the event, the SenseMart Go robotic store demonstrated how the OS can coordinate these components in a real-world retail setting.
Why it matters: SenseMart OS offers a glimpse of how embodied intelligence could move beyond demonstrations and into everyday commercial environments.
- The launch also reflects a broader debate over what will define the next phase of competition in embodied intelligence: robot hardware, data, models, or systems?
Why retail: SenseTime already had years of experience in the sector, with product-recognition systems processing millions of orders daily and a database covering more than 300,000 SKUs, or distinct product types and variants, Dr. Yi Shuai, co-founder and chief scientist of SenseTime Smart Retail, told TechNode.
- Unlike warehouses or factories, retail requires robots to interact with people, testing perception, interaction, decision-making, and execution at the same time.
- Retail also offers clear business metrics, such as sales, repeat purchases, and ROI, that can be used to evaluate whether the technology is working.
Zoom in: Dr. Yi said SenseMart OS focuses on three elements that support four core capabilities — perception, interaction, decision-making, and execution — as SenseTime looks to expand the system beyond retail into other offline service industries.
Business outcomes: The system targets metrics such as customer satisfaction, repeat purchases, and transaction completion, not just robot task success.
Multiple robot types: It can connect wheeled dual-arm robots, humanoids, grippers, and other hardware depending on the task.
- Retail data: It combines visual, transaction, inventory, and voice-interaction data across the customer journey to train retail agents.
By the numbers:
- One employee can currently support three to five stores, cutting average staffing to about 0.2–0.3 workers per location, Dr. Yi says.
- The robots can run for about six hours on a charge and typically need charging twice a day. If a customer arrives while one is charging, it can leave the dock to serve them first.
- Strong Shanghai locations can reportedly break even in about six to 12 months; average sites take around a year.
- Each store carries about 150–200 SKUs, roughly one-third to one-fifth the assortment of a small convenience store.
The scaling play:
- SenseMart OS can learn one new equipment-handling skill per week at a store and replicate it across existing and future locations.
- Training a new store model has fallen from two to three months at the first location to about one to two weeks today.
- Standardized deployments can begin operating the day after installation, according to Dr. Yi.
On the ground: SenseMart Go now operates more than 20 stores across cities including Shanghai, Hefei, Shenzhen, Qingdao, and Yancheng, spanning campuses, malls, offices, industrial parks, events, and public spaces.
What to watch: Speakers at the event pointed to different areas where the next wave of value in embodied intelligence could emerge.
- Industry know-how: “The industry’s greatest value will move toward both ends — upstream supply chains and downstream industry data and expertise,” Dr. Yi said. SenseTime Smart Retail is focusing on the latter through retail.
- Systems: “What matters more is getting embodied intelligence into production and daily life than the system or foundation model itself,” said Weisheng Xu, co-founder and CEO of Jiutian Yuanshu. For now, systems can help bridge gaps between foundation models and real-world deployment.
- Real-world data: “Every time a robot works in a real-world setting, the flywheel of models and data turns once,” said Min Yuheng, co-founder and CEO of Zerith Robotics. Real deployments, he argued, are key to improving both data and models.