Latest in AI
AGIBOT Worken and GO-3 Preview: What Changed at Innovation Day
AGIBOT wants to measure useful robot work, but the definition of a successful hour matters as much as the model behind it.
AGIBOT introduced Worken and GO-3 Preview at its October 29 Innovation Day in Shanghai. Worken proposes a way to measure productive robot output, while GO-3 Preview combines reasoning, prediction and action. For robotics buyers, the useful question is how model improvements translate into work that needs less supervision.
What is AGIBOT Worken?
In AGIBOT's announcement, Worken means effective intelligent work hours. Its formula multiplies active robots by autonomous working hours and an output-efficiency ratio measured against a task-specific reference.
The company calls this an evolving framework, not an established industry standard.
That distinction matters. A robot count is easy to compare, but it says little about how much work gets finished. A fleet that spends most of a shift waiting for someone to reset failed tasks should not be evaluated like a fleet that operates independently.
The harder problem sits inside the output-efficiency ratio. Buyers need to know what counts as acceptable output, who sets the reference and how interrupted tasks are treated. A single number becomes useful only when those definitions travel with it.
What GO-3 Preview changes
GO-3 Preview uses AGIBOT's WRAM, short for World Reasoning Action Model. AGIBOT reports leading results on RoboDojo, RoboTwin and RoboColiseum. These are company-presented claims, not an independent deployment audit.
For context, the RoboTwin 2.0 research project evaluates dual-arm manipulation and varies conditions such as lighting, clutter, tabletop height and instructions. Those variations help test whether a policy can handle changes beyond a carefully arranged scene.
A benchmark can help choose a model for further testing. It cannot, by itself, price the human effort needed to keep that model operating in a warehouse. The missing bridge is a deployment test that records interventions, completed tasks and recovery time together.
What is available now?
The event also launched G2 Max and Link-U OS 1.0. AGIBOT says the operating system is in selected-partner trials, with full open sourcing planned by the end of 2026. GO-3 remains explicitly a preview; the announcement provides no public API price or download instructions.
The existing Link-U OS repository describes an Alpha built around the Expedition A2 robot. Its documented foundation includes AimRT communication middleware, ROS 2 compatibility and tools connecting simulation with deployment. That public Alpha should not be confused with general availability of the newly announced 1.0 release.
For an engineering team, this is a reason to separate architecture evaluation from a production commitment. An existing repository can show how components fit together without establishing that the latest release is ready for an arbitrary robot or operating site.
The test robotics buyers should ask for
Consider two systems that finish the same number of acceptable handling tasks. If one requires frequent operator attention, its apparent hardware productivity hides a staffing cost. If the other works independently but needs long maintenance stops, looking only at autonomous hours hides a different cost.
Worken needs a cost and intervention record beside it. Neither a model score nor a productivity formula should erase the conditions under which the work happened.
The next useful evidence would be a repeatable customer-site evaluation with a defined task, acceptance criteria and a clear accounting of human assistance. That would make the gap between a stronger robot model and a better operating business much easier to inspect.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders turn complex AI launches into clear stories that developers and buyers can understand and share.