Why 2 of 6 Agibot Robots Failed Their Convenience Store Probation

Real-World Convenience Store Test

On November 2, Agibot placed six Lingxi G2 humanoid robots into an operating convenience store in Shanghai's Xuhui district for a continuous 24-hour live stream. Each robot had a dedicated role: Ahao handled greeting and inquiries, Huobao performed picking and restocking, Juanwang patrolled shelves, Kaoshao acted as barista, Dingjiang managed checkout and prepared food, and Lengjie cleaned tables. The store remained open to regular customers throughout peak and off-peak hours across four shifts.

Evaluation criteria mirrored real employment standards: on-duty duration, task completion volume, customer feedback, task success rate, response time, and anomaly-handling ability. Four robots met the bar; Lengjie and Juanwang fell short on engagement consistency and positional stability, so they did not "pass probation." The article notes that such unvarnished results are rare in an industry accustomed to polished demos, and the public test served as a stress test of Agibot's full technology stack under genuine commercial conditions.

Perception, Manipulation, and Scheduling Challenges in an Open Environment

The article identifies three layers of difficulty that a real convenience store imposes on embodied systems, contrasting them with controllable lab settings:

Fine-grained manipulation of deformable objects. Soft-packaged goods change shape during grasping, requiring real-time control adjustments based on force and tactile feedback. Handling fragile items like golden cakes demands precise force control — too little force drops the item, too much crushes it.

Long-horizon task planning and coupling. Restocking involves a chain of dependent subtasks: target recognition, inventory assessment, path planning, and state maintenance. If an item is occluded, the robot must first clear the obstruction. Parallel multi-tasking further requires sustained long-term state memory.

Obstacle avoidance in dynamic, open spaces. Customers and pets move unpredictably. The system must perceive and decide in real time under strict safety constraints, slowing down or aborting motions when necessary.

Lab environments offer calibrated lighting, known object poses, and scripted workflows (in-distribution conditions). A real store continuously generates out-of-distribution samples — a successful grasp once does not guarantee repeatability under new spatial conditions.

During the stream, invited "experience officers" deliberately increased difficulty: asking robots to distinguish between sugared and sugar-free cola, placing people and boxes in the robot's path, and shaking boxes while the robot attempted to grasp prepared food. The robots verbally complained "work is hard" but completed the recognition, detour, and adjusted-grasp sequences.

Genie Master: Closing the Deployment-Learning Loop

In open environments, the article argues, recovery resilience matters more than raw success rates. Agibot's Genie Master embodied agent framework acts as a shared "brain" for all six robots, handling task scheduling and linking execution outcomes to subsequent learning. Its core embodied execution protocol forms a "judge–execute–feedback" closed loop: the agent knows each robot's capabilities, issues execution guidance, receives real-time results, and uses them for the next judgment cycle.

This enables robots to self-diagnose failures and adapt strategies without human teleoperation or engineer debugging. In pre-deployment tests, Lingxi G2 equipped with Genie Master improved its process success rate on prepared-food packing, picking, and restocking from 28.6% on day one to 85.7% by day five, then stabilized.

Newly acquired skills are consolidated into reusable experience through two pathways: (1) successful and failed trajectories are summarized daily into long-term memory for direct recall on similar tasks; (2) effective adjustment processes become training data that refine motion policies. The combined flywheel means more robots and longer shifts yield more experience and faster improvement — "on the job, unlocking new skills."

This addresses a core bottleneck for large-scale embodied AI deployment: relying on human teleoperation and engineer debugging makes labor costs scale with robot count, breaking the business model. Genie Master shifts that burden to the robots themselves — execution results automatically feed back, policies automatically adjust. The convenience store data suggests this loop is initially viable.

Full-Stack Synergy as a Prerequisite for Physical AI

Breaking down the store tasks shows no single strong point can solve any problem alone. Precise grasping depends on hardware performance — joint precision, force control, hardware consistency. Fluid motion requires policy-control co-optimization. Reusable experience demands complete data recording and evaluation infrastructure. The article states that embodied AI's industry characteristic is that point breakthroughs cannot substitute for full-stack collaboration.

This philosophy underpins Agibot's ten-day release cadence starting October 22 (AI WAVE): AGIBOT WORLD 2026 dataset (4th edition), Genie Studio 2.0 work platform, heavy-payload G2 MAX, Genie Master agent framework, Genie Sim 3.5 simulation platform, GO-3 Preview foundation model, and Collective Experience Learning (CEL). The pieces are highly coupled: body stability determines data quality; data spans pre-training, post-training, and post-deployment continuous learning; model deployment relies on system integration, process adaptation, and multi-device coordination. Any weak link stalls the entire loop.

The 24-hour public convenience store test — placing the complete stack into a real store under public scrutiny — is itself a pressure test of that full-stack choice. The two robots that did not pass preserve the test's realism; showing both success and shortfall reveals exactly how far robots remain from job requirements. The purpose of open testing is to let technology face actual business validation and improve further. The 24-hour live stream has ended, but robots driven by real-world scenarios have just begun.