Most of the attention in AI still lands on chatbots. But the harder, and arguably more consequential, frontier is getting a model to move a physical body through the real world. On 29 October 2026, at its Innovation Day in Shanghai, the Chinese robotics company AgiBot (also styled AGIBOT, and known in Chinese as Zhiyuan Robotics) made two announcements that are worth separating from the usual humanoid-robot theatre. One is a model, GO-3 Preview. The other is a measurement proposal, Worken. Together they tell you how at least one serious player thinks the robotics race will be won, and judged.
What AgiBot announced
GO-3 Preview is what the field calls an embodied foundation model: a single large model that takes in what a robot sees and an instruction in plain language, and outputs the physical actions to carry it out. It is the successor to AgiBot’s earlier GO-1 line. The company describes its architecture as WRAM, a World Reasoning Action Model, which chains three steps that older action models tended to collapse into one: reasoning about the task, predicting the likely outcome of a move, and then generating the action itself. The “prediction” step is the interesting part, because it imports an idea from world models, letting the system anticipate what will happen before it commits to a motion.
The headline design goal is continuous learning across three stages: pre-training, post-training and deployment. Pre-training mixes two data sources: first-person human videos, which show how people handle objects, and robot trajectories, which tie what a camera sees to the physical action and its result. For post-training, AgiBot built a distributed system it calls Genie Evolver, which combines roughly 100 physical robots with physics simulation and world models so robots can practise, grade their own attempts and improve by repetition. The third stage, learning during deployment, is the ambitious one: a robot that keeps getting better at a task on the job rather than freezing at the capability it shipped with.
The demonstrations were chosen to show generalisation and dexterity. AgiBot showed GO-3 cooking jianbing, a Chinese savoury crepe that needs several coordinated delicate movements; assembling Lego after fine-tuning, with an agent sequencing the steps; and, in a learning demo, beating the average human score at tabletop curling after four hours of online practice, and pulling five blocks in a row from a Jenga-style tower without toppling it. The company says the model can handle dexterous manipulation without a separate fine-tune for every new activity, and claims first place on three embodied-AI benchmarks: RoboDojo, RoboTwin and RoboColiseum. AgiBot plans to release the complete model as open source by the end of 2026. The same event also brought a new hardware platform (AHA3), an omni-modal interaction model (WITA-Omni), an industrial robot (G2 Max) and an embodied-intelligence operating system, Link-U OS 1.0, also slated to go open source by year’s end.
Worken: counting work, not robots
The second announcement is a quieter but sharper idea. AgiBot proposed Worken, a framework for measuring the productive work robots actually deliver once deployed. The name fuses “work” and “token”, and the company frames it as “effective intelligent work hours”. The proposed formula is deliberately simple:
Worken = Number of Active Robots × Autonomous Working Hours × Output Efficiency Ratio
The last term compares what a robot completes against a reference benchmark for a task, for example the acceptable output of a human worker under established standards. The point, in the words of co-founder and CTO Peng Zhihui, is to “move beyond measuring how many robots are built or delivered and begin measuring how effectively they can perform useful work, autonomously and consistently, in the real world.” AgiBot says the framework is still in development and will be refined with deployment data.
That reframing matters because the industry’s loudest number, units shipped, is close to a vanity metric. Two fleets of the same size can deliver wildly different value depending on how often a human has to step in. Worken is, in effect, an attempt to make utilisation and autonomy the scoreboard. It is the robotics version of a shift software teams know well: from counting lines of code to measuring what actually ships and works.
How new is any of this, really
Embodied foundation models are now a crowded field. Google DeepMind’s Gemini Robotics line, Physical Intelligence’s models, Nvidia’s GR00T, Figure’s Helix and AgiBot’s own GO-1 all pursue the same vision-language-action recipe. What GO-3 claims as differentiators are the explicit reasoning-and-prediction loop and learning during deployment. Both are plausible directions rather than solved problems. The world-model idea it leans on is the same one driving systems like Genie 4, and the open-source posture echoes other recent moves to open the robotics stack.
The scepticism writes itself. The benchmark wins and the oft-quoted jump in task success from about 29% to 86% over five days of market tests are company-stated and not independently verified. RoboDojo, RoboTwin and RoboColiseum are largely simulation benchmarks, and success in simulation has a long history of not surviving contact with a real kitchen. The demos are curated; cooking one jianbing on stage is not the same as cooking a hundred unattended. And the model is a “Preview”: the full weights and the open-source release are a promise for “the end of 2026”, not something a developer can download today. Even AgiBot’s claim to be the biggest humanoid maker rests on first-half-2026 shipment estimates that the trackers disagree on, with IDC putting it above 8,600 units and roughly a third of the market while Counterpoint reads it higher.
Worken deserves its own caveat. A productivity metric proposed by the current market-share leader, with a subjective “output efficiency ratio” at its core and no independent standards body behind it, can easily become a marketing instrument rather than a neutral yardstick. The idea is good; who defines and audits it will decide whether it means anything.
Why it matters for the future of IT
Strip away the hype and two durable signals remain. First, the serious robotics bet is now on generalist models that learn continuously, not on hand-coded skills, which is the same architectural bet that reshaped software with large language models, now aimed at the physical world. Second, a leading vendor is trying to change the success metric from installed base to delivered work, a sign the industry expects to be judged soon on economics, not demos. Both arrive against a politicised backdrop: the United States moved in mid-2026 to restrict new Chinese-made humanoids on security grounds, which makes a Chinese champion open-sourcing its model a competitive and geopolitical move as much as a technical one. For developers, the thing to watch is not the stage demo but whether the open-source GO-3, when it lands, actually lets third parties reproduce the claims on their own hardware.
Sources
- Interesting Engineering: Humanoid robot achieves 85% success rate with AGIBOT’s GO-3 model
- PR Newswire / RoboticsTomorrow: AGIBOT Introduces Worken to Measure Productive Robot Work at AGIBOT Innovation Day 2026
- The AI Insider: AgiBot Introduces Worken Metric to Measure Productive Robot Labor
- Pandaily: AgiBot Unveils GO-3 Preview Robot Model, Tops Three Leaderboards and Plans Open Source
- Robotics and Automation News: Agibot claims top spot in global humanoid robot shipments in first half of 2026
- AGIBOT: IDC Data Shows AGIBOT Led Global Humanoid Robot Shipments in H1 2026