What happened
AGIBOT announced GO-3 Preview as its next-generation embodied foundation model. The company says the model is built around WRAM, short for World Reasoning Action Model. In AGIBOT's description, WRAM links three functions: reasoning about a task, predicting what may happen, and selecting an action. The stated objective is to help a robot generalize beyond task sequences it has already learned.
The company demonstrated what it described as zero-shot and few-shot capabilities. It also described a training process spanning pre-training, post-training, and continuing improvement after deployment. AGIBOT reported leading results on RoboDojo, RoboTwin, and RoboColiseum, including first place for task success rate and process score on RoboDojo. These are company-reported benchmark claims; the source explicitly says Humanoids Daily did not independently verify the comparisons.
The open-stack commitment is as important as the model preview. AGIBOT co-president Yao Maoqing said GO-3 Preview and the Genie Master embodied-agent framework will be fully open sourced by the end of 2026. The company also said Link-U OS 1.0, its robot operating system, is being trialed by selected partners and is intended for full open sourcing by year end. AGIBOT says the Link-U project will be donated to the OpenAtom Foundation. Those statements describe a roadmap, not a presently downloadable or broadly accessible release.
The announcements sit alongside new data-collection and hardware components. MEgo Glove is a wearable intended to capture human manipulation demonstrations as training data. Air and Ultra versions capture finger poses; Ultra additionally captures three-dimensional fingertip and palm force data and torque information. AGIBOT also introduced AHA3, its third-generation humanoid hardware architecture, plus updated manipulation hardware and an industrial G2 Max robot. AGIBOT lists G2 Max with dual-arm handling capacity of up to 40 kg and a working-height range from 0.1 to 2.5 meters.
Finally, AGIBOT proposed Worken, or “Effective Intelligent Work Hours,” as a deployment measure. Its formula is active robots multiplied by autonomous working hours and an output-efficiency ratio. The company distinguishes active task-performing machines from inventory, autonomous work from merely being powered on, and acceptable output from undifferentiated operating time. It presents Worken as an evolving, task-specific framework rather than an established standard.
What the archive adds
This is a new event, but it belongs beside RoboParty's RP1 full-stack open-source humanoid release. RP1 framed openness across hardware, firmware, deployment, and learning. AGIBOT's announcement makes a comparable full-stack claim at the roadmap level, centered on an embodied model, an agent framework, and an operating system.
Trend signal
The Worken proposal supports the direction described in Agent economics moves to completed work: the useful unit is not a model call or machine count, but verified work delivered. AGIBOT has not established an industry standard, however; it has offered a formula whose reference benchmark and acceptable output still need to be defined for each task.
Why now
The source frames the event around deployment rather than robot shipment counts. Its premise is that factories, warehouses, and service environments need a way to distinguish fleets with the same number of robots but different intervention rates and acceptable output. AGIBOT connects model and software improvement to extracting more useful work over a machine's operating life.
Connections to other events
The earlier RP1 Field Note is the closest related event because it also treated an open humanoid stack as more than a single model or component. AGIBOT's proposed stack differs in status: RP1 was presented as a released GPL-3.0 project, while GO-3, Genie Master, and Link-U OS are promised for full open sourcing by the end of 2026.
Why it matters
Embodied-AI teams often have to assemble training data, a control or agent layer, runtime software, and physical hardware across separate suppliers. A genuinely open release covering GO-3, Genie Master, and Link-U OS could give teams more of that path to inspect, adapt, and integrate. The MEgo Glove announcement also points to a practical bottleneck: manipulation data with pose and force information, not only language or vision data.
The immediate implication is more cautious. Builders cannot yet evaluate the license, weights, interfaces, hardware support, security boundaries, or performance of the promised stack. Treat this as an architecture signal and a watch item, not as a deployable alternative to existing robotics tooling.
What happens next
- AGIBOT publishes a concrete release scope for GO-3, Genie Master, or Link-U OS. (3 months, 60%). Watch for: repositories, licenses, model weights, or API and hardware compatibility documentation.
- External teams test the reported embodied-model results or reproduce them on related benchmarks. (6 months, 45%). Watch for: independent RoboDojo, RoboTwin, or RoboColiseum evaluations.
Limits and uncertainty
The public source is based on materials supplied by AGIBOT and labels the item paid content. It does not provide model size, weights, licensing terms, training-data details, inference requirements, safety controls, supported robots, or a public release date more precise than the end of 2026. The source reports AGIBOT's benchmark positions but says they were not independently verified. Worken is a proposal, and its output-efficiency ratio depends on task-specific reference standards that AGIBOT says still require refinement with operational data.
Skill to build now
Define a deployment scorecard for one physical workflow before choosing a robotics model. Separate the number of active machines, minutes of genuinely autonomous operation, intervention frequency, and the share of outputs that meet an explicit acceptance criterion. That makes a proposed metric such as Worken testable against the measures that actually matter to an operator.
Reader takeaway
AGIBOT's substantive news is a planned open embodied-AI stack, not a model release that builders can use today. Its accompanying Worken formula is a useful reminder: evaluate robots by accepted autonomous work, while waiting for the code, weights, licenses, and independent tests needed to judge the stack itself.
Open question
Will AGIBOT's year-end open-source commitment include sufficiently permissive licenses, usable hardware interfaces, and reproducible evaluations for outside robotics teams to adopt the stack?