The Pulse
AGIBOT Pairs GO-3 Preview With a New Measure of Robot Work
AGIBOT introduced Worken, a proposed measure of productive robot work, alongside its GO-3 Preview model at an event in Shanghai. The company says the framework combines active robots, autonomous working hours and task-specific output effici
AI.info Team · Written with AI · fact-checked ·
AGIBOT’s proposed Worken measure multiplies the number of active robots by their autonomous working hours and an output-efficiency ratio. The formula aims to count productive work, not just machines produced, delivered or left powered on. The company introduced it alongside GO-3 Preview, its next-generation embodied AI model, at an event in Shanghai on October 29.
Worken is a framework under development, not an established industry standard. Its usefulness will depend on how companies define acceptable output, record human intervention and compare performance across different tasks. AGIBOT says it intends to refine the measure using operational data from real deployments.
Worken counts task output, not powered-on time
AGIBOT calls Worken “Effective Intelligent Work Hours.” Its three parts are active robots, time spent working autonomously and an output-efficiency ratio measured against a reference for a specific task. The company suggests that reference could be the amount of acceptable work completed by a human under established operating standards.
The distinction changes what a deployment figure can tell a buyer. Two fleets of the same size might deliver different amounts of useful work if one needs more human assistance or completes fewer acceptable tasks. Worken is meant to account for those differences, though the release does not set out a common method for measuring them.
AGIBOT co-founder, President and CTO Peng Zhihui said the measure is intended to shift attention from production totals to performance in use, referring to how much productive work robots deliver.
“Worken is our attempt to develop a new way of answering that question.”
— Peng Zhihui, Co-founder, President and CTO of AGIBOT
GO-3 Preview links reasoning, prediction and action
AGIBOT describes GO-3 Preview as an embodied foundation model built around WRAM, short for World Reasoning Action Model. The company says the framework combines reasoning, prediction and action so a robot can anticipate the consequences of a movement, adjust to changes in its surroundings and handle tasks beyond previously learned sequences.
At the event, AGIBOT demonstrated what it described as zero-shot and few-shot capabilities. It also presented a training approach called Physical Intelligence Scaling, combining pre-training and post-training with continued improvement after deployment. The stated aim is for robots to build on structured training and experience from operating in the physical world.
The company said GO-3 Preview ranked first in task success rate and process score on RoboDojo, achieved the highest overall score on RoboTwin, and ranked first in overall score on RoboColiseum. AGIBOT’s release provides rankings but no underlying scores or detailed comparison conditions, so the results describe the company’s account of its evaluations rather than an independently verified comparison.
Factory and service tasks pose different tests
AGIBOT presented Worken as relevant to industrial handling, logistics, inspection and commercial services. The company pointed to existing robot deployments in tablet production, semiconductor packaging and testing, CNC machine tending and logistics, as well as visitor services at Guangzhou Metro and the Chimelong tourism complex.
Those settings do not share one simple definition of success. A factory task can be judged by completed handling work and cycle time; a public-facing service may depend on whether a robot understands a request and responds appropriately. AGIBOT says Worken’s output ratio should use task-specific benchmarks, but the release does not provide examples of measured Worken results from those sites.
A proposed measure meets a preview model
The announcements sit alongside AGIBOT’s AHA3 hardware platform, which the company says combines foundations for locomotion, interaction and task execution. AGIBOT also launched its G2 Max industrial robot and Link-U OS 1.0, while describing the latter as being trialled by selected partners and slated for open-source release by the end of 2026.
Worken could make robot productivity easier to discuss only if operators can agree on what counts as autonomous time, acceptable output and necessary human help. GO-3’s reported benchmark rankings offer one kind of evidence about model performance; Worken is an attempt to describe what that performance produces on the job.