AGIBOT Launches Worken to Measure Productive Robots

Humanoid makers have spent years counting units, demonstrations, and factory pilots. AGIBOT wants the industry to count productive hours instead. At its Shanghai Innovation Day on October 29, 2026, the company introduced Worken, a proposed measure of useful autonomous robot work, alongside the GO-3 Preview foundation model, the AHA3 hardware platform, and the heavy-duty G2 Max robot. The package signals a shift from proving that a humanoid can move to proving that it can perform repeatable jobs without constant human rescue. That is a harder claim because deployment exposes every weak link at once: perception, planning, manipulation, safety, uptime, service, and economics. Worken tries to give customers and investors a common language for that reality. The metric could improve accountability if AGIBOT publishes a strict definition and comparable data. It could also become another marketing score if vendors count easy hours, exclude interventions, or report workloads that do not match paid industrial work.

What Actually Happened

AGIBOT introduced Worken at its annual Innovation Day in Shanghai on October 29, 2026. The company combines the words "work" and "token" to describe a unit intended to measure productive robot activity in real environments. The official AGIBOT announcement says the framework is designed to move attention beyond the number of robots built or deployed toward how much useful work they perform autonomously and consistently. Co-founder, president, and chief technology officer Peng Zhihui presented the concept as part of a wider deployment strategy. AGIBOT did not publish a complete public formula in the announcement, so questions remain about how Worken handles task difficulty, human intervention, failure, safety stops, and idle time. Even so, the choice of metric is revealing. A manufacturer that asks to be judged on productive work is accepting a tougher standard than shipment volume, especially when many humanoids still spend more time in demonstrations and supervised trials than in continuous operations.

The event also introduced the AHA3 integrated hardware platform, the GO-3 Preview robot foundation model, the WITA-Omni 1.0 model, Link-U OS 1.0, and the G2 Max industrial robot. The Humanoids Daily event report places Worken inside a full-stack release that includes data, training, motion intelligence, and deployment tools. G2 Max targets heavier work and is rated to carry 40 kilograms, according to coverage of the announcement. AGIBOT also showed a liquid-cooled joint system, new dexterous hands, and its MEgo data-collection glove. That breadth matters because productive hours depend on more than a foundation model. A robot needs durable joints, task-specific end effectors, safe payload handling, reliable software updates, and a way for workers to teach new motions. AGIBOT is presenting those pieces as one operating system for commercial deployment rather than isolated research projects.

AGIBOT attached scale claims to the technical releases. A China AI Wire report from the event says Peng reported 20,000 general-purpose AI robots produced in the first half of 2026 and nearly RMB 2 billion in revenue. Those figures make the Worken proposal more than a theoretical benchmark because the company has a fleet large enough to collect deployment data across varied settings. The official release highlights uses in industrial manufacturing, logistics, retail, and service environments, though it does not provide audited utilization or autonomy rates by category. The event date and product list are also confirmed by RobotToday's Innovation Day briefing. The next step is disclosure. Production and revenue show commercial reach, but they do not show how many hours the robots work, how often people intervene, or whether customers expand after the first deployment.

Why This Matters More Than People Think

Robot shipments are a weak proxy for economic value. A unit can leave a factory, enter a showroom, and count as deployed even if it performs only a few scripted demonstrations each day. Productive time exposes the difference between inventory and labor capacity. If Worken counts hours only when a robot completes a defined task within quality and safety limits, it can help buyers compare automation options on output rather than spectacle. A warehouse operator could compare robot hours with human-assisted hours, throughput, rework, and downtime. A manufacturer could track whether a humanoid performs several jobs or remains tied to one station. The metric would also change sales conversations. Instead of promising a general-purpose worker, a vendor would commit to a measurable amount of completed work. That can support outcome-based pricing, service-level agreements, and financing. A robot with a higher purchase price may still win if it produces more verified work per shift, needs fewer interventions, and can switch tasks without weeks of integration.

The metric also addresses the industry's demo problem. Humanoid videos compress hours of setup into a short clip and rarely show failed attempts, battery swaps, teleoperation, or recovery after an unexpected object appears. A standardized productivity measure could force vendors to report the denominator. For example, 100 autonomous task hours means little unless readers know the scheduled time, number of interventions, task success threshold, and conditions. Worken could combine those factors, but AGIBOT must resist the urge to hide complexity inside one score. Customers need both the headline number and the underlying components. The strongest model would report productive minutes, task quality, human support minutes, energy use, safety events, and availability. That resembles manufacturing's established measures such as overall equipment effectiveness, which separates availability, performance, and quality. Humanoids need a comparable discipline because their flexibility makes simple cycle-time comparisons incomplete. A robot that changes jobs quickly may create value even if its peak speed trails a fixed industrial arm.

However, critics argue that a vendor-defined metric can favor the vendor's own products and workloads. AGIBOT could choose tasks that fit its hardware, count supervised work as autonomous, or weight easy repetitions the same as complex manipulation. Competitors might publish incompatible versions that prevent comparison. The risk is largest when a single score becomes a substitute for raw operational data. An honest Worken framework needs independent governance, public definitions, and audit rules. It should state exactly when the clock starts, what ends a productive interval, how interventions reduce credit, and how quality failures are counted. It should also prevent double counting when several robots cooperate. Customers must be able to reproduce the number from logs, not accept a dashboard controlled by the supplier. AGIBOT can make Worken credible by opening the specification and inviting other vendors, integrators, insurers, and industrial users to challenge it. Without that process, the name may spread while the measurement remains too flexible to support purchasing decisions.

The Competitive Landscape

AGIBOT competes with a fast-growing group of humanoid and embodied-AI companies. Tesla is moving Optimus toward factory production, Figure is developing its next hardware generation and industrial deployments, Unitree sells lower-cost humanoids and robot dogs, and UBTECH has placed Walker robots with Chinese manufacturers. Agility Robotics focuses Digit on logistics, while Apptronik and Boston Dynamics pursue industrial work through different hardware designs. Chinese peers also include Galbot, Fourier, and Xpeng. Each company can report units, pilot sites, tasks, or hours in a way that emphasizes its strongest result. Worken is an attempt to set the scoreboard before the market settles on one. AGIBOT's broad portfolio, reported production scale, and domestic supply network give it many data points. Tesla can use its own factories, Figure has named automotive experience, and Agility has a focused logistics use case. A common productivity measure would make those strategies easier to compare, which is precisely why competitors may resist AGIBOT's definition.

The historical parallel is cloud computing's shift from server ownership to metered consumption. Enterprise technology once focused on how many servers a company bought and how powerful each machine was. Cloud providers changed the unit of value to compute consumed over time, then built pricing, monitoring, and service guarantees around that measure. Worken imagines a similar transition for physical labor. Buyers would care less about robot count and more about verified work delivered. The comparison has limits because physical tasks vary more than processor cycles. Moving one identical box is not equivalent to sorting mixed returns, assisting a person, or handling a fragile component. Safety and space also shape the value of each hour. Still, metering can unlock new business models. Vendors could lease robots by productive hour, financiers could underwrite fleets against utilization, and customers could expand capacity without buying every unit. AGIBOT's challenge is to define a unit stable enough for contracts while preserving the task context that makes the number economically useful.

Foundation models create another competitive layer. GO-3 Preview is AGIBOT's attempt to connect pretraining, practice, and learning during deployment. Google DeepMind, Nvidia, Physical Intelligence, Skild AI, and several Chinese labs are building models that map vision and language to robot actions. Nvidia supplies simulation, chips, and reference models to many manufacturers, which may let it become the neutral software layer. A hardware vendor such as AGIBOT can optimize model, joints, hands, and data collection together, but customers may fear lock-in. A model provider can support many robot bodies, though cross-platform performance is difficult because sensors, kinematics, and control loops differ. Worken could become a bridge between these layers if it measures task output regardless of the model or body. It could also reinforce AGIBOT's stack if only its products expose the required logs. The market will test whether buyers prefer one accountable supplier or a modular system that lets them replace the model, hardware, and integrator separately.

Hidden Insight: Robot Accounting Becomes the Product

The non-obvious value of Worken is accounting. Companies cannot scale a new labor asset until they can measure it, price it, insure it, and assign responsibility when it fails. A humanoid fleet creates operational questions that unit shipments cannot answer. Finance teams need depreciation schedules and utilization. Operations teams need uptime and task quality. Safety teams need near-miss and intervention records. Vendors need evidence for maintenance and warranty claims. An auditable productive-work unit can link those records. Once every task carries a start time, completion test, autonomy level, energy cost, and human-support cost, the robot becomes legible to the rest of the business. That data can be more valuable than a polished demonstration because it supports budget decisions. AGIBOT may be using Worken to make its fleet financeable and manageable, not merely to win a messaging contest. The company that defines the operational ledger could influence how customers buy robots even when several manufacturers supply the hardware.

That ledger could support outcomes-based contracts. A customer may hesitate to buy a humanoid because the useful life, maintenance burden, and task flexibility are uncertain. A vendor that charges for verified productive work absorbs more of that risk. The customer pays when the robot completes agreed tasks, while the vendor manages hardware, software, and service. Worken could define the billable unit, much as cloud usage meters define a compute invoice. This model rewards reliability because downtime reduces vendor revenue. It also gives AGIBOT a reason to improve models after deployment rather than treat the hardware sale as the finish line. Yet the contract needs careful boundaries. A customer controls floor layout, upstream materials, network availability, and worker interaction. The vendor controls the robot and much of its software. Worken records must distinguish failures caused by the robot from blocked work caused by the site. Otherwise both sides will dispute the bill, and the metric will add friction instead of reducing it.

Productivity data can also improve the foundation model. Each recorded task contains a natural training signal: the instruction, observed environment, action sequence, intervention, and final quality result. A fleet of 20,000 robots could generate a large stream of examples that laboratory datasets lack. AGIBOT's MEgo glove and simulation tools add demonstrations before deployment, while Worken can label which behaviors produced useful results afterward. That creates a feedback loop between commercial work and model training. The benefit depends on data quality. Repeated easy tasks can dominate the dataset and teach the model to optimize common motions while neglecting rare failures. Customer sites may also restrict video and telemetry because the data reveals processes, products, or worker identities. AGIBOT needs privacy controls, on-site processing, and contractual clarity about model training rights. If it solves those issues, productive-work logs can become a defensible advantage that grows with fleet usage rather than a static score shown at an annual event.

The bear case is that Worken cannot normalize different jobs without losing the detail that matters. One hour of carrying standardized totes is not equal to one hour of manipulating cables, serving customers, or working beside people. Weighting tasks by difficulty invites subjective choices and gaming. Leaving them unweighted makes cross-industry totals misleading. The solution may be a family of measures rather than one universal token. AGIBOT could publish raw productive time, task-specific throughput, quality, autonomy, and human-support ratios, then let customers assign economic value. Worken would function as a data standard, not a magical exchange rate for labor. That narrower ambition is still useful. Manufacturing adopted multiple measures because no single number captures a plant. Humanoid operators will likely do the same. AGIBOT should avoid claiming that all robot work is fungible and instead make each score traceable to a defined task. Credibility will come from detail, even if detail makes the headline less simple.

What to Watch Next

Within 30 days, watch for AGIBOT to publish a technical definition of Worken. The document should explain the unit, autonomy threshold, treatment of teleoperation, quality checks, downtime, safety stops, and human intervention. It should include worked examples across at least two settings, such as logistics and assembly, because a metric tested on one workflow cannot support a broad claim. The company should also clarify GO-3 access, supported robot bodies, and whether customers can inspect task logs. G2 Max specifications need detail beyond the 40-kilogram payload, including reach, runtime, speed under load, safety certification, and planned delivery dates. Developers will look for Link-U OS documentation and interfaces that connect third-party tools. If AGIBOT releases only promotional summaries, Worken will remain a concept. A public specification and downloadable sample logs would show that the company expects customers and competitors to test the measure rather than repeat its name.

Within 90 days, the key evidence is a deployment report with denominators. AGIBOT should name a customer, define the task, state scheduled hours, productive hours, interventions, failures, and output quality. A before-and-after comparison against manual work or fixed automation would help buyers understand the economic value. Watch whether a third party audits the results and whether another robot maker experiments with the metric. Adoption outside AGIBOT would make Worken more credible, even if a competitor proposes changes. Fleet reliability will matter as G2 Max and AHA3 move from the stage to work sites. Heavy payloads increase the importance of braking, thermal management, collision avoidance, and service intervals. The new liquid-cooled joints could improve sustained work, but field data must show lower heat-related downtime or higher duty cycles. Revenue claims should also be separated by hardware, software, services, and related businesses so investors can see whether productive deployments are driving repeat demand.

Within 180 days, look for contracts priced or guaranteed using productive work. A customer paying by verified hour would show that Worken has moved from marketing into commercial operations. Another marker is fleet expansion at existing sites. Repeat orders indicate that robots create value after the pilot team leaves. AGIBOT should report the share of work completed autonomously, the median intervention interval, and whether software updates raise output on deployed hardware. Competitors may answer with their own utilization standards or join an industry group to define a neutral framework. Regulators and insurers will also shape the measure by asking for safety events and operator responsibility. The strongest outcome is not AGIBOT owning a proprietary token. It is a market where robot vendors compete on transparent completed work and customers can audit every claim. If Worken becomes that shared language, the October 29 launch may matter more than any single robot shown on the stage.

A robot shipment measures inventory. Productive autonomous hours measure whether the machine earns its place on the floor.

---