AgiBot Debuts GO-3 Preview and WITA-OMNI, Plus a Worken Unit for Robot Output

At this year's AgiBot Innovation Day, co-founder, president and CTO Peng Zhihui introduced Worken, a new unit for physical AI intended to measure a robot's effective intelligent working hours. The formula on stage multiplies the number of active robots by autonomous working time and by an output efficiency ratio.

The three factors map to deployment scale, whether a robot can keep working without human takeover, and how much it actually produces. The intent is straightforward: the language-model era uses tokens to measure usage and cost, and physical AI needs a commercial anchor tied to real work so competition shifts from demo polish to sustained productivity.

Two new models underpin that measurement. GO-3 Preview, a foundation model aimed at manipulation, ranked first on three embodied-AI benchmarks: RoboDojo, RoboTwin and RoboColiseum. It uses an embodied-native architecture called WRAM, short for World Reasoning Action Model, which places task understanding, reasoning, future-state prediction and action generation in one learning framework.

AgiBot says GO-3 Preview emphasizes joint scaling of models and data during pretraining, improving transfer and adaptation in zero-shot and few-shot tasks. After fine-tuning, robots running the model can grasp, align and insert parts, and complete assembly tasks under agent orchestration. Swapped onto the bipedal Lingxi X2, it can open doors with both hands and pick up objects while kneeling, demonstrating cross-embodiment transfer.

Post-training and deployment carry the capability into real sites. AgiBot built Genie Evolver, which combines more than a hundred real robots, physical simulation and world models for post-training. Once on the job, the Genie Master agent framework handles task planning and error recovery, turning on-site mistakes into experience for later runs.

The second model targets interaction. WITA-OMNI, an end-to-end omni-modal model, took first place on Daily-Omni, TimeLens-Bench, EmbSpatialBench and MMAU. It models speech, vision and text inputs together with speech, facial expression and motion outputs in a single system, so a robot can listen, watch, speak and move at the same time.

Notably, WITA-OMNI needs no wake word. It identifies who is speaking by combining sound direction, voiceprint, face and lip movement, then uses gaze and tone to judge when a response fits, even while the robot is busy with another action. The design is a fast-slow split: an Interaction Engine in the foreground answers immediately, while a Background Agent handles reasoning, planning, tool calls and long tasks.

At the system level, AgiBot's Lingqu OS targets deterministic latency, holding end-to-end jitter below 4 microseconds under 1000Hz task loads, and packages system services as reusable Skills that developers can call directly. AgiBot plans to fully open-source Lingqu OS 1.0 by the end of this year and donate the project to the OpenAtom Foundation, alongside 200 million lines of production-validated code, more than 5,000 ecosystem packages and an end-to-end toolchain.

Deployment numbers are the test of the bundle. A 3C production line built with ODM partner Longcheer began scaled production in March 2026 and had run stably for more than seven months by Innovation Day; during a six-day live stream, robots completed over 64,824 operations and built 17,625 tablets with a 99.99% success rate at a 22-second cycle. In retail, 100 Lingxi X2 units at an Aishida store drove 2.23 million yuan in sales over three days, while nearly 500 robots cover more than 100 roles at Chimelong's Hengqin space park. The open question is whether Worken becomes a shared industry measure or stays an internal KPI.