AGIBOT has completed a 24-hour convenience store robotics trial in Shanghai, where six G2 wheeled humanoid robots collectively performed 1,730 tasks involving customer assistance, coffee preparation, product restocking and cleaning. The experiment tested how embodied AI systems handle continuous operations, changing environments and unexpected customer-facing situations, offering a practical assessment of autonomous robotics beyond controlled demonstrations.
Humanoid robots are increasingly moving from research laboratories and controlled demonstrations into environments where they must interact with people, handle physical objects and respond to unpredictable situations. AGIBOT is testing that transition through a 24-hour convenience store trial in Shanghai, using six G2 wheeled humanoid robots to perform everyday retail operations.
The trial took place at an operating convenience store in Shanghai’s Xuhui District from 10 a.m. on November 2 to 10 a.m. on November 3. Across the 24-hour period, the six robots completed a combined 1,730 tasks spanning customer service, product handling, food preparation, shelf management and cleaning.
Four robots passed the final performance assessment, while two failed to meet all evaluation criteria.
The event concluded AGIBOT’s AI WAVE technology showcase, which began on October 22. The initiative brought together the company’s robotics hardware, simulation technologies, AI models, operational software and embodied AI capabilities in a real-world retail environment.
The trial was also livestreamed across online platforms, attracting more than two million viewers, according to AGIBOT.
Six Robots Take on Continuous Retail Operations
The six G2 robots were assigned different responsibilities to simulate the operational demands of a convenience store.
Their activities included customer assistance, shelf inspections, product restocking, coffee preparation, checkout support, hot food service and table cleaning.
AGIBOT reported the following task totals:
- 683 customer-service interactions.
- 369 product pickups and deliveries.
- 147 coffee orders.
- 126 breakfast servings.
- 277 shelf inspection and restocking tasks.
- 128 table-cleaning tasks.
The robots operated across four distinct periods: lunchtime, evening peak hours, overnight operations and the following morning’s breakfast rush.
This scheduling exposed the systems to different operational conditions, including changing customer traffic, product availability and task requirements.
The assessment considered time on duty, task completion, customer feedback, response times and the ability to handle unexpected situations.
Practical evaluations also tested whether the robots could distinguish between visually similar products, return rearranged merchandise to its correct shelf position and navigate around obstacles blocking an aisle.
These tests highlight a central challenge for autonomous robotics: completing a programmed movement is different from understanding what needs to happen when the surrounding environment changes.
Why Retail Is a Difficult Test for Embodied AI
Convenience stores may appear relatively simple compared with industrial environments, but their constantly changing conditions make them useful testing grounds for embodied AI.
Products can be moved, packaging can vary, customers can interrupt robot movements and objects may appear in unexpected positions.
A robot picking up a bag of candy, for example, must identify its orientation and grip it without disturbing nearby products. Preparing food requires additional control over movement and gripping force, particularly when handling delicate items.
During the trial, AGIBOT said its G2 robots used real-time visual information to identify misplaced merchandise and return products to designated shelf positions.
One robot detected gaps on shelves and replenished products without receiving a specific instruction. Another distinguished spills from countertop patterns and selected appropriate cleaning tools.
These activities require several AI capabilities to work together: visual perception, object recognition, spatial understanding, task planning and physical manipulation.
This is where embodied AI differs from conventional generative AI applications. While large language models primarily process and generate information, embodied AI connects perception and decision-making with physical actions through robotic hardware.
The retail environment therefore provides an opportunity to examine whether AI systems can translate their understanding of a situation into reliable physical execution.
Genie Master Connects Planning, Execution and Learning
A central component of the trial was AGIBOT’s Genie Master embodied AI agent framework.
The system is designed to coordinate robotic tasks, monitor execution and use operational feedback to improve subsequent performance.
Through its Embodied Execution Protocol (EEP), Genie Master connects task planning, physical execution and feedback analysis.
The framework organizes tasks according to individual robot capabilities and operating conditions. Robots then execute assigned activities using action policies and real-time observations, with the resulting execution information returned to Genie Master for analysis and troubleshooting.
This creates a feedback loop between task assignment and physical performance.
For example, if a robot fails to grasp an item correctly or encounters an unexpected obstacle, the execution results can be reviewed to identify the problem and adjust subsequent actions.
AGIBOT says the framework can retain relevant experiences in long-term memory and use successful adjustments as training data to refine robotic action policies.
This approach is important because autonomous robots cannot depend entirely on fixed instructions when operating in dynamic environments. They need systems that can respond to changing circumstances and build on previous operational experience.
The company has also used quieter overnight periods to provide additional practice for G2 robots performing tasks requiring precise manipulation.
However, the trial results do not independently establish how much performance improved through this learning process or whether the improvements will transfer consistently to different retail environments.
Physical AI Moves Toward Real-World Deployment
The experiment reflects a broader shift in artificial intelligence development toward physical AI systems capable of performing work in the real world.
Technology companies are increasingly investing in robotics platforms that combine machine learning, simulation, computer vision, autonomous navigation and AI agents.
NVIDIA, for example, has expanded its robotics ecosystem through accelerated computing, simulation technologies and physical AI development tools. Other robotics developers are pursuing different combinations of foundation models, hardware platforms and autonomous control systems.
AGIBOT’s approach emphasizes the connection between robot execution and continuous operational feedback.
The commercial challenge, however, extends beyond completing a large number of tasks during a demonstration. Retail operators need robots that can perform consistently across extended shifts, handle different products, interact safely with customers and operate at an acceptable cost.
The fact that four of the six robots passed the final assessment also illustrates that reliable autonomous operation remains a development challenge.
AGIBOT plans to continue testing its robots across additional roles and real-world environments, using operational feedback to improve reliability, adaptability and task execution.
For the robotics industry, the Shanghai experiment provides another example of how embodied AI is being evaluated against practical operating requirements rather than isolated laboratory benchmarks.
The next milestone will be determining whether these capabilities can be replicated across different stores, workflows and operating conditions. That will help establish whether humanoid robots can move from technology showcases toward scalable commercial applications.
Market Landscape
The embodied AI market is developing around the convergence of foundation models, computer vision, robotics hardware, simulation and autonomous control systems.
Unlike software-only AI applications, physical AI must translate model decisions into actions through motors, sensors and robotic manipulators. This introduces additional requirements involving precision, latency, safety and reliability.
Retail is an important testing environment because it combines repetitive tasks with frequent changes in object positions, customer behavior and operating conditions.
AGIBOT’s Genie Master framework reflects another emerging trend: integrating AI agents with robotic execution systems so that operational feedback can inform subsequent task planning and model improvement.
The competitive landscape includes NVIDIA’s robotics ecosystem and specialized humanoid and autonomous robotics developers. Long-term differentiation will depend on task reliability, manipulation accuracy, adaptability, deployment economics and the ability to transfer learned capabilities between environments.
Top Insights
- 24-hour operational testing: Six AGIBOT G2 robots completed 1,730 combined tasks across customer service, food preparation, restocking and cleaning.
- Real-world adaptability: The trial tested visual recognition, object manipulation, obstacle avoidance and responses to changing retail conditions.
- AI agent coordination: Genie Master connects task planning, robotic execution, feedback analysis and experience-based improvements.
- Performance remains a challenge: Four of six robots passed the final assessment, highlighting the continuing difficulty of reliable autonomous operations.
- Commercial deployment is the next test: AGIBOT plans additional trials to evaluate whether capabilities can transfer across different environments and responsibilities.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
