Zhengxing Innovation launches retail physical AI 24/7 service solution at APRCE 2026

The solution pairs robot hardware built for shelf restocking, store inspection and related transport work with a unified robot management platform that handles task allocation, status monitoring and manual takeover. According to the company, it closes the loop from task acquisition and environmental perception through movement, recognition and grasping to task execution, and can be deployed without modifying store shelves or customer flow paths, an approach it calls zero-modification deployment.

Two hardware products were introduced at the exhibition: H1, a general-purpose bipedal humanoid robot for spaces shared with people, and C1, a wheel-arm transport and restocking robot designed for narrow aisles. The company says the robots can take over hauling, restocking, inventory counting and inspection, freeing store staff for other service work, and can absorb repeated physical labor at night so that employees do not have to stay awake on overnight shifts.

Zhengxing Innovation said it will work with a major global chain retailer to test and commercialize the products, validating the system's performance in real, open retail environments while refining robot task flows and operating systems, verifying return on investment and extending use to hot food preparation, inventory management, delivery sorting and packing. Commercial service is planned for 2027, sold either through direct purchase or a Robotics-as-a-Service subscription covering hardware, software, operations and maintenance and ongoing upgrades.

The company was co-founded by serial entrepreneurs Yao Song and Yang Yuxin, Tsinghua University young scholar Yu Chao, and the multinational industrial group CP Group. In a keynote titled Physical AI and the Future of 24/7 Retail, founder and CEO Yao Song said the company chose to start from retail so that robots could work and evolve in the most complex and most live commercial environment, and to become a globally trusted robotics service provider.

Zhengxing Innovation argues that retail stores are among the most commercially valuable settings for embodied intelligence but also among the hardest to enter: large numbers of stores, many SKUs, dense shelving, narrow aisles and constantly shifting customer flow and displays demand far more perception, real-time decision-making and adaptation than other sites. Most robots in the industry, it says, work in structured, fixed settings such as back-of-house warehouses and closed stockrooms and struggle in crowded, complex front-of-store environments. The company says its robots can complete 99 percent of tasks autonomously.

H1 uses a soft exterior design, supports interaction in more than 20 languages and can display 45 expressions and 36 gestures, with an open API for secondary development. C1 carries a narrow-body omnidirectional chassis that moves through convenience-store aisles without shelf changes; its working height spans 100mm to more than 1,750mm to cover full shelf ranges, and it performs transport, restocking, loading and unloading in tight spaces. Multiple sensors let it detect pedestrians, with emergency braking under 10cm, and it carries front and rear depth cameras plus 360-degree panoramic vision. M1, the management platform, is described as a self-developed hardware-software operating base for multiple robot types and cross-scenario work, connecting device onboarding, task execution and data review for large-scale deployment, scheduling and continuous optimization of collaborative tasks. The company plans heavier-load versions and an A1 precision-operation robot for checkout and bagging, forming a standard, dexterous and heavy-load product line.

The company attributes this capability to its embodied brain, a physical intelligence system combining embodied models, algorithms, data and infrastructure that can keep iterating while running real tasks. Its self-developed lightweight world-action model SLM-0.5 handles the relationship between environment and action and reached a 98.6 percent average task success rate on the LIBERO benchmark, with inference 24 times faster than mainstream alternatives, the company said. A reinforcement-learning post-training system called STEAM, built on RLinf, uses light task data and human feedback for rapid retail adaptation and reached 95 percent success in a validation task restocking designated products; results and corrections feed back into operating policies in a loop of execution, feedback, learning and redeployment. Rpent, an agent foundation co-initiated with Tsinghua University and Infinigence, combines general model task understanding and planning with VLA specialist models for fine manipulation, plus memory, tools and robot interfaces. Given an instruction such as restocking a row of shelves, Rpent breaks down the goal, plans the order of execution and dispatches navigation, recognition and manipulation modules, monitoring progress and re-planning when items are misplaced or a grasp fails.

The company says it is accelerating replication and deployment of the solution with support from the retail business ecosystem of partners including CP Group. QbitAI published the material provided by Zhengxing Innovation, and the views are those of the original author.

Editor's Summary

Zhengxing Innovation released a Physical AI system for human-robot collaboration in retail at APRCE 2026, comprising H1 humanoid and C1 wheel-arm robots plus an M1 management platform, with commercial service planned for 2027 through direct sales or RaaS subscriptions. The company says its robots complete 99 percent of tasks autonomously and cites benchmark results for its SLM-0.5 world-action model and STEAM training system. It plans tests with a global chain retailer and later expansion into hot food, inventory management, delivery sorting and packing.