AI-generated conceptual illustration of robot work
A batch of robots completing delivery is a business milestone for the manufacturer. For the factory receiving the equipment, the next questions are more specific: Which processes can they handle, how many qualified operations can they complete in a day, and how many people are needed to supervise them?
On October 29, AgiBot showcased the GO-3 Preview embodied foundation model at its 2026 Innovation Day in Shanghai, officially launched the G2 Max industrial robot and Link-U OS 1.0, and proposed a robot productivity measurement framework called Worken. Together, these announcements point to how robots perform after entering the site.
Powered On for Eight Hours Does Not Necessarily Mean Eight Hours Worked
According to the definition published by AgiBot, Worken is composed of the number of robots actually participating in tasks, autonomous operating time, and output efficiency. Among these, output efficiency must be compared against a reference benchmark for a specific task, examining the amount of work that meets established standards.
This approach separates some numbers that are easily conflated: equipment being manufactured does not mean it has already taken a job; having taken a job does not mean it operates autonomously continuously; and completing an action does not necessarily produce qualified output.
At a parts loading and unloading workstation, the difference is intuitive. After the robot picks up a part, it still needs to place it correctly into the fixture, match the cycle time of downstream equipment, and detect placement failures. If it frequently waits for manual adjustment, the actual output for the entire shift may differ greatly from the speed shown in a single demonstration.
This is a work example used to understand the evaluation method. It also shows that before counting effective work, task boundaries must first be defined: whether manual replenishment counts as auxiliary time, how retries are recorded, and what counts as completing one qualified operation. With different definitions, the same number may express completely different operating levels.
Worken is currently a framework proposed by AgiBot that still needs to be refined through deployment data, and it cannot yet be regarded as a unified industry standard. Its reference value depends on whether task definitions and recording methods are clear enough for users to review.
GO-3 Needs to Answer Constantly Changing On-Site Questions
The GO-3 Preview showcased this time adopts an architecture AgiBot calls WRAM, combining reasoning, prediction, and action. The company hopes this will help robots anticipate the consequences of actions, adapt to environmental changes, and handle situations beyond existing task sequences.
These capabilities have clear test targets on site. Changes in the position of incoming materials, objects occluding one another, and fixture offsets will all require the system to reassess; when downstream equipment pauses, the robot may need to wait or stop operating.
The value of model upgrades can be observed along these changes: which manual interventions it reduces, whether it can detect execution failures, and how long it takes to resume operation. Only by recording anomalies and their handling process together can the performance of continuous operation be explained.
Productivity evaluation ultimately has to return to cost. For the same qualified output, if more maintenance, integration, or downtime handling is required, the deployment return will change. Enterprises need to put operation records together with full costs to determine whether a robot is worth continuing to use and where the next batch of equipment should be deployed.
AgiBot's announcement this time puts models, equipment, and work measurement into the same discussion. The most convincing material to follow will be long-term operating data accumulated at specific workstations: how tasks are completed, how anomalies are handled, and how inputs are converted into qualified output.