AGIBOT Reports 1,730 Tasks in G2 Store Trial

Hero photograph: AGIBOT / PR Newswire, an archival G2 museum demonstration from October 2025, not this store trial. Original release; editorial-use permission. Original image retained; the site may crop it responsively.

The store trial, counted carefully

AGIBOT's breakdown: 683 customer interactions, 369 pickups/deliveries, 147 coffee orders, 126 breakfast servings, 277 shelf checks/restocking jobs and 128 table cleanups. Evaluators tested similar products, displaced gum and an obstructed aisle. Assessment covered duty time, completion, feedback, response and unexpected conditions. Attempts, thresholds, supervision, energy management and commercial terms remain undisclosed; no retailer is named. These are company claims.

The breakdown describes a store with several kinds of work, rather than a robot repeatedly moving one object between fixed points. Product retrieval requires the right selection as well as a stable grasp. Restocking adds a destination constraint: an item has to return to a useful shelf position. Cleaning introduces another decision, because the robot must choose an action appropriate to the surface and the mess.

These jobs also interact. A robot approaching a shelf might need to yield to another machine carrying a drink, then resume without losing track of its item. That makes recovery from interruptions a central retail capability. The reported pass count assesses whole robots; it cannot be converted into a success rate for individual tasks.

Under the hood: hardware versus service performance

G2 is a wheeled mobile manipulator, with arms mounted above a moving base. AGIBOT's July 30 user manual, which the company describes as dynamically updated, provides the specifications below. They describe the product; they are not measurements from this trial.

Degrees of freedom count independently controlled motions. More articulation can offer alternative reaching positions, but the useful measure at a shelf is whether the complete system places the right object correctly. Arm capacity alone cannot establish that: object recognition, approach angle, grip and release must all work together.

The manual lists an optional charging dock and teleoperation that can time-share an operator across robots. It also specifies only 15 mm of obstacle clearance. An event lasting a day cannot establish uninterrupted autonomy or the uptime of each machine. A battery specification describes an energy constraint, not the replenishment strategy chosen for a particular event.

The dimensions make store layout part of the engineering problem. A passage wide enough for navigation may still be a poor place to stop and reach across a shelf. Floor transitions, temporary stock boxes and access to cleaning equipment need attention before adding another machine. Those integration questions remain relevant even when the arms can lift the merchandise comfortably.

Independent measurement research helps explain the distinction. NIST's mobility research program, updated in April 2026, addresses localization under disturbances including moving obstacles, wheel slip and changing loads. Its evaluation agenda includes adaptability, efficiency and completion quality. That work does not assess G2; it shows why a specification sheet needs to be supplemented by tests of the integrated machine in its intended surroundings.

AGIBOT's software case

AGIBOT attributes the trial's coordination to Genie Master. In its separate October 29 GO-3 Preview announcement, the company described Genie Master as an agent framework linking task planning, execution feedback and retained experience. Its Embodied Execution Protocol, or EEP, defines interfaces for available capabilities, guidance, feedback and unfinished work. AGIBOT describes two learning paths: preserving useful experience for later decisions and using execution records to improve action policies, the software that chooses robot movements.

That architecture offers a plausible way to organize a store workflow. Consider a hypothetical delivery: fetching the object, reaching the destination and confirming the handoff are separate steps. If the handoff is interrupted, a coordinator needs to know what remains to be done. Otherwise, restarting the whole request could duplicate work, while marking it complete could leave the customer empty-handed.

AGIBOT also offers tools for the less visible work of deployment. Its October 26 Genie Studio 2.0 release describes three modules. Flywheel connects data collection, training, simulation and deployment. Agent lets teams assemble skills and decision logic into visual workflows. Monitor provides fleet status, task progress, fault alerts and remote intervention. The company says fault records and resolutions can feed back into workflow and model updates.

Those tools could help an operator locate a stalled step rather than merely see that a robot has stopped. The distinction matters when multiple machines share a coffee station or a narrow route: fixing the arm's motion would not resolve a scheduling conflict. These are available capabilities and possible applications, not a verified account of the software configuration used in Shanghai.

The October 29 announcement discussed an earlier store-learning demonstration. Its claims about autonomous improvement cannot be carried forward as evidence of the supervision used during the later six-robot event. Software capability and the operating arrangement of a particular trial need separate documentation.

What a store has to integrate

A workable installation needs more than a path between shelves. The store must decide where a robot waits, where it hands over goods and how a customer can safely decline or change a request. A waiting position that blocks a popular product could undermine an otherwise successful delivery. These are design questions for the whole service, not just the motion controller.

Shared equipment creates another bottleneck. A hypothetical coffee workflow might involve reaching the machine, collecting a cup, transporting it and finding the recipient. If the pickup point is occupied, the system needs an orderly way to wait or redirect work. Adding robots would not necessarily increase throughput if every job still depends on the same constrained station.

Food and cleaning duties also require clearly defined boundaries. Before extending a trial, an operator would need to establish which objects each robot handles, how tools are prepared and cleaned, and which exceptions go to staff. None of those operating arrangements can be inferred from a count of completed servings or cleanups.

Human help can be part of a viable service. Its frequency and duration still belong in the cost calculation, especially if several machines request help during a rush. A NIST-authored mobile-manipulator survey separates completion, time, quality and quantity, and distinguishes teleoperation from autonomy. It did not evaluate G2; the useful lesson here is to examine the complete service and its support burden together.

The next test: a different store

A second venue would expose different counter heights, shelf spacing, floor transitions and customer routes. It could test how much of the application transfers and how much has to be rebuilt. Changing product packaging would add a useful recognition challenge without changing the underlying job of fetching or restocking merchandise.

The strongest comparison would preserve the task definitions while reporting accepted completions, retries and assistance by job. Peak periods should have their own results, because a quiet overnight interval cannot answer whether the system keeps up when several customers need service at once. Records of charging, resets and unavailable time would show how much of the shift was usable.

These are proposed evaluation steps, not announced commitments. A named retailer's acceptance criteria and repeated use would add commercial evidence. The aim should be to establish a reproducible installation that fits an ordinary store's operating constraints, rather than a larger headline total from one venue.

Frequently asked questions

From activity to dependable service

The most revealing retail challenges occur at the connections between jobs. Movement, manipulation, customer interaction and recovery all compete for time and attention. A convincing next result would show how those connections behave under sustained demand, with the full support burden visible.