1. Context and Key Points

In November 2026, a convenience store located in the Jing’an district, Shanghai, hosted a 24-hour field test with six AGIBOT G2 robots. During the period, the robots executed 1,730 tasks that included customer service, product replenishment, payment management, and inventory control. At the end of the test, four of the six robots exceeded the performance evaluation criteria established by the store operator, while the remaining two showed limitations in multilingual interaction and adaptation to unexpected customer flows.

The experiment is relevant because it constitutes one of the first large-scale deployments of service robots with dialogue capability based on large language models (LLMs) within a real retail environment. The results offer an empirical reference for Chief Technology Officers (CTOs) and operations managers evaluating the feasibility of replacing or complementing human staff with autonomous physical agents.

2. Technical Highlights

The AGIBOT G2 is the second generation of the service robot family developed by the Chinese company AGI Robotics. The architecture combines an omnidirectional mobile platform with a 7-degree-of-freedom arm joint, an RGB-D camera, 2D LiDAR sensors, and an onboard processing unit based on NVIDIA Jetson AGX Orin. The choice of the Orin allows for the execution of language model inferences of several hundred million parameters without relying on 5G connectivity, which reduces the latency perceived by the customer to less than 300 ms in simple dialogues.

At the software level, the robot employs a three-layer AI stack: (1) perception, which fuses vision and LiDAR data using a lightweight convolutional network trained with distillation techniques; (2) planning, which uses a reinforcement model based on PPO to generate safe trajectories in real time; and (3) dialogue, which relies on an LLM aligned via DPO (Direct Preference Optimization) and fine-tuned with retail interaction data in Mandarin and English. The preference alignment was performed with a set of 150,000 conversation turns labeled by human operators, which ensures coherent and respectful responses in sales contexts. The retraining process of the dialogue model was executed on the Alibaba Cloud, where weights were updated weekly through federated learning that incorporates the interaction logs from the test. This approach allows the robots to adapt their repertoire of responses to local consumption patterns without exposing sensitive data. From an integration standpoint, the AGIBOT G2 connects to the store management system (POS) via a RESTful API that allows for checking prices, availability, and promotions in real time. The messaging architecture is based on MQTT with QoS 1, ensuring the delivery of critical commands even under intermittent network conditions. Regarding physical safety, the robot incorporates a collision detection module based on a stereoscopic vision algorithm that stops locomotion when identifying obstacles at less than 0.3 m. Additionally, the arm features a torque controller that limits force to 15 N, preventing damage to fragile products or users. The two robots that did not pass the evaluation presented recurring failures in managing customer queues when demand exceeded 30 simultaneous customers. Subsequent analysis revealed that the dialogue model's token cache capacity (approx. 4k tokens) was saturated, causing truncated responses. This bottleneck suggests that future versions will need to expand the context window or implement history compression techniques.

3. Sectoral Impact and Market Repercussions

The convenience retail sector in China is experiencing increasing pressure to reduce operating costs and improve the customer experience. According to data from the China Chain Store & Franchise Association, the average gross margin in convenience stores is around 22% and staff turnover exceeds 30% annually. The introduction of service robots like the AGIBOT G2 offers a way to stabilize the workforce and free human employees from repetitive tasks.

From a supply chain perspective, the robot's ability to execute automatic replenishments reduces lead time by an estimated 15%, by eliminating the reliance on manual intervention to move products from the warehouse to the shelf. This saving translates into greater availability of critical SKUs and, potentially, an increase in sales of up to 3% in high-turnover categories. In the capital market, investors have shown growing interest in service robotics startups. Since the beginning of 2025, venture capital destined for Chinese robotics companies has exceeded 1.2 billion dollars, with a particular focus on solutions that integrate LLMs and computer vision. The partial success of the AGIBOT G2 test reinforces the narrative that the convergence of generative AI and edge hardware is reaching commercial maturity. However, mass adoption faces regulatory barriers related to data privacy and civil liability in the event of an accident. The regulations of the National Development and Reform Commission (NDRC) require that AI systems interacting with consumers maintain conversation logs for at least 30 days and that manufacturers implement decision auditing mechanisms. In terms of competitiveness, the main international players (Boston Dynamics, SoftBank Robotics) are developing versions of service robots with similar dialogue capabilities, but AGI Robotics' advantage lies in native integration with local payment platforms and the availability of training data in Mandarin, which reduces localization friction.

4. Market Outlook

Analysts from the research firm IDC China point out that the 66.7% pass rate (four out of six robots) is sufficient to justify expanded pilots in store chains with more than 500 points of sale. They recommend that operators prioritize the segmentation of high-value tasks (contactless payment, returns management) for the robots, while human staff focus on consultative sales and complex incident management.

From the perspective of AI architecture, experts in federated learning warn that the collection of interaction data in retail environments must be accompanied by robust anonymization protocols to comply with the Personal Information Protection Law (PIPL). The lack of such protocols could expose operators to fines of up to 5% of their annual turnover. Operations managers of convenience chains that have already implemented delivery robots in internal warehouses observe a 12% reduction in direct labor costs. However, they stress that the integration of legacy systems (ERP, WMS) represents the main technical challenge, as the AGIBOT G2 API needs custom adapters for each platform. Regarding sustainability, the energy consumption of the AGIBOT G2 is estimated at 150 W in active mode, which is equivalent to 3.6 kWh in a 24-hour day. Compared to the average consumption of a full-time employee (approx. 1 kWh per office hour), the robot presents a competitive energy footprint, provided that the energy comes from renewable sources. Investment strategists suggest that private equity funds should consider the acquisition of startups that offer fleet management software layers, as the ability to remotely monitor and update hundreds of robots will be a key differentiator in the next five years.

5. Next Steps

In the next 12 months, AGI Robotics plans to launch a “G2-Plus” version that incorporates an 8k token context window and a multilingual voice recognition module that supports Cantonese and Japanese. It is expected that the improvement in context capacity will eliminate the bottlenecks observed in the Shanghai test, raising the pass rate to more than 85%.

In the medium term (18, 24 months), the company plans to integrate a route planning engine based on dynamic graphs that allows robots to simultaneously optimize product replenishment and customer service, reducing the average service time from 45 s to less than 30 s per interaction. By 2028, analysts project that at least 20% of convenience stores in major Chinese cities will employ at least one service robot in their daily operations, driven by the reduction in hardware costs (the list price of the AGIBOT G2 is expected to fall by 30% compared to 2026) and by the maturity of AI regulatory frameworks.

6. Conclusion and Assessment

The deployment of six AGIBOT G2s in a Shanghai convenience store demonstrates that service robotics with dialogue capability is ready to move from the pilot phase to commercial adoption. The results indicate that, with a well-aligned AI architecture and careful integration with store management systems, robots can execute hundreds of daily tasks with a success rate of over 60%.

For CTOs and operations managers, the message is clear: the priority must be (1) to establish a data infrastructure that guarantees privacy and traceability; (2) to select high-return use cases where automation reduces operating costs without compromising the customer experience; and (3) to plan a hardware and software upgrade roadmap that includes expanding the context window and multilingual expansion. Only through a comprehensive strategy will it be possible to transform the 24-hour test into a sustainable competitive advantage.