AgiBot Tops Seven Robotics Benchmarks With Two New Models, and Wants Worken to Be the Token of Physical AI

Chinese robotics firm AgiBot (Zhiyuan) used its 2026 Innovation Day at Shanghai's World Expo Center on October 29 to make an argument that the humanoid industry has matured past demos: what it needs now is a unit of account. Co-founder, president and CTO Peng Zhihui — better known online as Zhihui Jun — unveiled "Worken," a metric the company defines as effective intelligent work hours, calculated as the number of active robots multiplied by autonomous working hours multiplied by an output efficiency ratio.

The framing is deliberate. Steam power was measured in horsepower, electricity in kilowatt-hours, digital AI in tokens. Worken is AgiBot's bid to be the token of physical AI — a way to bill, compare and benchmark robots by what they actually accomplish rather than by spec sheets or deployment counts. Peng's own line to reporters: "Ten thousand units off the line is a victory for manufacturing; ten thousand units on the job is a victory for physical AI."

The company backed the metric with two new models it says top the leaderboards. GO-3 Preview, an embodied foundation model built on a World Reasoning Action Model (WRAM) architecture that fuses task understanding, reasoning, future-state prediction and action generation, claimed first place on three embodied-AI benchmarks: RoboDojo, RoboTwin and RoboColiseum. All rankings are company-reported and not independently audited. The model supports zero-shot and few-shot task transfer and crosses embodiments — mounted on the bipedal X2 humanoid, it opens doors and kneels to pick things up. A post-training system called Genie Evolver combines roughly a hundred real robots, physics simulation and world models, while a Genie Master agent framework handles task planning and recovery from mistakes in deployment.

On the interaction side, WITA-OMNI 1.0 is a full-modality model that jointly handles listening, vision, language, facial expression and gesture. AgiBot says it tops four benchmarks — Daily-Omni, TimeLens-Bench, EmbSpatialBench and MMAU — and its claimed response time averages around 600 milliseconds with a fastest of about 400, close to natural human conversational latency. There is no wake word: the model fuses sound direction, voiceprint, face and lip movement to work out who is talking to it and when to respond. It understands more than 100 languages and speaks 29 plus seven dialects, by the company's count, and can run fully offline. A fast-slow dual-system design keeps a front-end interaction engine responding instantly while a background agent handles reasoning and long tasks.

The hardware announcements rounded out the stack. The third-generation AHA3 architecture collapses the three "intelligences" — motion, interaction, task — onto a single SoC, which the company says cuts physical volume by 50 percent and improves response speed by 90 percent, both vendor figures. The new G2 Max heavy-duty robot carries 20 kilograms on one arm with a 50-kilogram peak dual-arm load, seven-degree-of-freedom force-controlled arms and hot-swappable batteries. And Lingqu OS 1.0, described as the first embodied-AI-native operating system, supports 15 categories and more than 100 peripherals with end-to-end latency jitter under 4 microseconds at 1,000 Hz workloads. It will be fully open-sourced by year-end and donated to China's OpenAtom Foundation, with 200 million lines of production-validated code and 5,000-plus ecosystem packages to follow.

The commercial numbers, all company-reported, were the quiet headline: AgiBot says its 20,000th general-purpose AI robot rolled off the line in September, that first-half 2026 shipments reached 9,700 units with revenue approaching 2 billion yuan (about 280 million dollars), and that it holds more than 43 percent of the global humanoid robot market. Its Longqi 3C production line — billed as the first fully covered by embodied intelligence — has run for over seven months; in a six-day livestream the robots performed 64,824 operations, produced 17,625 tablets at a 99.99 percent success rate and matched skilled human workers at a 22-second cycle time. Nearly 500 robots work at Hengqin Chimelong theme park across more than 100 roles.

Whether Worken becomes an industry standard or a marketing construct depends on whether third parties adopt the formula — and on whether customers agree to be billed by it. But the direction of travel is clear across the industry: after years of unit shipments and park demos, the humanoid sector is converging on the same question enterprise software answered decades ago. Not "how many did you sell," but "how much useful work did it do while nobody was watching."