Zhi Yuan used Worken to find out exactly how much work the robot had done

Once the robots leave the showroom, the number of units delivered is just the beginning of the question. Buyers are more concerned with how long these devices can operate autonomously each day, whether the results they produce meet standards, and how many people are needed to step in and fill gaps. Worken, introduced by Zhiyuan at Innovation Day on October 29, 2026, integrates these questions into a single metrics framework, aiming to shift the discussion from the number of machines to actual output.

The term “Worken” is a combination of “Work” and “Token,” and Zhiyuan describes it as “effective smart working hours.” It is currently a framework proposed by enterprises that still needs to be refined through real-world deployment; it is not yet an established industry standard. Its value lies primarily in reframing the question, rather than serving as a “free pass” for the commercialization of robots.Official Announcement of the Zhiyuan Innovation Day

Three Factors That Separate Delivery from Work

The expression proposed by Zhiyuan is:Worken = Number of Active Robots × Autonomous Operating Time × Output Efficiency Ratio. Note that the number of active units does not include equipment that has merely been manufactured or is held in inventory; independent labor hours do not refer to machine runtime; and the output efficiency ratio must be compared to the benchmark for合格 output under established standards for a specific task.

This framework reveals a difference that is easily obscured by delivery figures: among two teams of robots with the same number of units, if one frequently has to wait for human intervention to handle anomalies while the other can continuously complete tasks that meet quality standards, their economic value may differ significantly. Counting only the number of units places the two on equal footing; only by factoring in autonomous operation time and output conditions can the gap be identified.

However, when actually implementing this, it is important to avoid being misled by averages. Since different machines have varying operating times and efficiencies, one cannot simply multiply the team’s average operating time by its average efficiency. A more prudent approach is to maintain records for each individual machine and task, and then aggregate the data using a consistent methodology. Otherwise, a machine that is highly efficient but operates infrequently might have an disproportionately large impact on the average.

The most difficult aspect to standardize is the definition of acceptable output.

Carrying a box of goods and handling a tourist inquiry cannot be directly compared based solely on duration. For the former, outcomes can be defined in terms of weight, distance, cycle time, and damage rate; for the latter, one must assess the accuracy of the response, whether the issue was resolved, and when to escalate the matter to a human agent. Even if ten tasks are completed, their difficulty, quality, and business value may vary.

Therefore, for Worken to become a useful language for procurement, the key is not to assign a single, attractive number to all tasks, but to enable fair comparisons among similar tasks. Buyers and sellers must, at a minimum, agree in advance on task boundaries, input conditions, qualification criteria, and standard operating procedures. If the benchmark is too simplistic, the relative efficiency of bots can be exaggerated; counting only successful segments also hides the costs of preparation and failure.

Human intervention requires particular transparency. Whether a remote operator takes over during an anomaly, on-site personnel reposition materials, or an engineer modifies a process—all of these actions may help the equipment complete its task, but they do not equate to the robot operating completely autonomously. Autonomous operating hours and hours requiring human assistance should be recorded separately, so that customers can determine whether performance improvements stem from the machine’s capabilities or from increased human input. This is a necessary step for the metrics framework to produce comparable data.

New hardware and models must be evaluated using the same criteria.

This Innovation Day also showcased the AHA3 hardware platform and the GO-3 Preview embodied model, and announced the G2 Max and Link-U OS 1.0. These respectively cover hardware integration, task capabilities, industrial operations, and system fundamentals. GO-3 was still presented as a Preview; Link-U OS 1.0 was being tested by select ecosystem partners at the time. Zhiyuan plans to fully open-source the platform by the end of 2026; this plan should not be interpreted as a completed open-source release.Product and Open Roadmap Overview

From Worken’s perspective, whether these updates are important should be judged based on concrete work outcomes. If improved heat dissipation reduces downtime during continuous operation, it may increase effective working hours; if a model can handle more anomalies, it may reduce the need for manual intervention; and if a unified software interface reduces the effort required for deployment across different environments, it may lower implementation costs. These are verifiable causal pathways, rather than business benefits that automatically materialize upon the release of a technology.

Similarly, benchmark results from models cannot be directly translated into factory production capacity. Task completion in a test environment is not equivalent to the actual conditions on the factory floor, which involve fluctuations in material supply, equipment wear and tear, conveyor congestion, and safety requirements. What is more useful to customers is a complete record of continuous operation under agreed-upon conditions, rather than a few demonstrations showing peak performance.

Effective working hours still need to be considered in conjunction with costs

Even if Worken improves, it does not necessarily mean the project will be more cost-effective. If, in an effort to get more work done, costs related to equipment maintenance, computing power, energy consumption, or on-site support rise even faster, the cost per unit of effective output may still increase. The purchasing party needs to calculate not only the costs of equipment and deployment but also ongoing operations and maintenance expenses, as well as the personnel allocated for handover and incident handling.

A more practical approach to pilot testing is to first select a task with a high repetition rate and easily verifiable results, record the cost and quality of the original operation, and then have the robot perform the task under the same conditions. After the pilot concludes, evaluate not only the quality of the output but also downtime, instances of human intervention, and recovery times. While the conclusions drawn from this approach may not be as grand as those from a “general-purpose workforce” pilot, they will provide a clear answer as to whether it is worth scaling up the deployment.

What makes Worken most noteworthy is that it shifts the focus from “whether a robot can perform a task” to “how many acceptable results it can consistently produce on its own.” Whether this can become a standard adopted by more customers depends on whether task benchmarks can be made public, whether raw data can be traced, and whether different manufacturers can agree on a common framework. Formulas are just the starting point; the real challenge lies in reliably quantifying actual work.