Spectators at robotics showcases have grown accustomed to elaborate stunts and synchronized dances, but both were largely absent at Agibot’s Innovation Day event on October 29. Instead, a wall at the Shanghai Expo Center, where the event was held, displayed research papers highlighting the company’s recent advances.
The emphasis reflected a challenge facing the robotics industry. Over the past three years, robots have learned to run, jump, and perform backflips, but these isolated technical achievements do not necessarily translate into reliable performance in complex environments. As artificial intelligence models such as GPT-6 Astra demonstrate the ability to plan physical tasks, deploying robots in real workplaces has become a more pressing test.
Agibot’s response covers much of the technology needed to put robots to work. Go-3 Preview handles task execution, AGILE addresses movement, and WITA-Omni handles interaction across different types of input. Its third-generation AHA3 hardware architecture and Link-U OS 1.0 are intended to connect those capabilities. Worken, a proposed measure of effective intelligent work hours, would assess the useful labor robots produce.
The announcements put sustained work, rather than demonstrations, at the center of Agibot’s next phase.
At the event, co-founder, president, and CTO Peng Zhihui said Agibot’s revenue approached RMB 2 billion in the first half of 2026 and that cumulative production of its general-purpose embodied robots had reached 20,000 units.
“Getting 10,000 robots off the production line is a victory for manufacturing. Getting 10,000 robots into jobs is a victory for physical AI,” he said in an interview afterward.
Models for tasks, movement, and interaction
A routine factory task illustrates the difficulty: walk to a shelf, pick up an item, and deliver it to a designated location.
A robot must locate the item and work out how to grasp it. While carrying it, the robot may need to avoid someone who suddenly walks past. If a worker changes the task, it must understand the new instruction.
Task execution, movement, and interaction all have to work. Agibot’s Go-3 Preview, AGILE, and WITA-Omni address those requirements.
Go-3 Preview is designed to help a robot understand the consequences of an action before taking it.
Vision-language-action, or VLA, models have become an important approach to embodied intelligence, connecting perception and instructions with actions. But knowing what to do next does not necessarily mean understanding what will happen. A robot reaching for a cup, for instance, could knock over something beside it.
Agibot describes Go-3 Preview’s decision-making process as “predict first, act second.” The model combines multiple camera views with instructions to establish the task and the surroundings, considers possible outcomes internally, then selects an action.
Agibot is also pursuing models that handle different forms of interaction from input to response, an area it sees as essential but comparatively neglected by major foundation model developers.
“Interaction is the necessary entry point, the first entry point, for a robot to create value. Every application depends on it,” said Xiong Yan, Agibot partner and co-president, who also heads central R&D and the interaction business.
Traditional voice interaction often requires users to say a wake word, ask a question, and wait for a reply. At the event, Xiong demonstrated a different approach with an A3 robot running WITA-Omni.
Without a wake word, the robot identified whether someone was addressing it amid the venue’s noise. It also handled interruptions and changes of question. According to 36Kr, its fastest on-site response took 400 milliseconds, and the model supports more than 20 forms of emotional expression and understands 73 languages.
Agibot also presented benchmark results. It said Go-3 Preview ranked first on RoboDojo, RoboTwin, and RoboColiseum, with a full open-source release planned by the end of 2026.
The company said WITA-Omni led four benchmarks: Daily-Omni for all-modality understanding, TimeLens-Bench for temporal localization in video, EmbSpatial-Bench for embodied spatial reasoning, and MMAU for complex audio understanding.
AGILE, Agibot’s framework for motion intelligence, addresses the physical response. It is intended to connect what a robot sees with how its body reacts, allowing movement to adapt to changing conditions.
Agibot is developing task execution, interaction, and movement together, but all three must coordinate in real time. They need a shared timeline and a consistent representation of the robot’s state. Otherwise, an instruction may not reach the motion controller correctly, or interaction may fail while the robot is moving.
Connecting models, hardware, and software
AHA3, Agibot’s third-generation hardware architecture, is designed as a unified control platform for motion, interaction, and task intelligence.
It addresses a familiar integration problem: separate models and control systems connected through cables and protocols can leave robots crowded with controllers, bulky, heavy, and difficult to cool.
Agibot said AHA3 halves the control system’s volume and nearly doubles its response speed.
Hardware integration alone does not solve the software problem. Sensors, controllers, and actuators from different suppliers use different interfaces and communication methods. If developers must write substantial low-level code for every new device, both development and upgrades become cumbersome.
Link-U OS 1.0 is intended to connect hardware, communications, and intelligent applications through a common operating system. Agibot said it is the industry’s first operating system built natively for embodied intelligence.
Agibot said its hardware interface standards let developers connect more than 100 peripherals across 15 categories, including cameras, sensors, and robotic arms, without having to handle each driver separately.
It added that, when a robot sends 1,000 motion commands per second, variation in the interval between commands stays below four microseconds. The aim is to keep rapid movement precisely timed.
Agibot announced a partnership with the OpenAtom Foundation to build an open-source community for humanoid robots in China and donate its Link-U OS project to the foundation.
The collaboration covers robot hardware, data and evaluation, operating systems, high-level intelligence, and motion control. It will include joint development, testing, validation, technical standards, and training for embodied AI specialists, with the aim of building an open global ecosystem.
Agibot plans to make Link-U OS 1.0 fully open source by the end of 2026. According to 36Kr, that would release approximately 200 million lines of code validated in mass production and more than 5,000 software packages.
The operating system packages capabilities for motion, interaction, and task execution so developers can reuse them instead of rebuilding each layer.
“You can never imagine what surprises partners will bring. That is the appeal of an ecosystem,” said Chuang Wang, an Agibot partner, senior vice president, and president of its general-purpose humanoid business. “We have to build the foundation well and leave more opportunities to our partners.”
How Worken measures robot productivity
For conventional industrial robots, the return on investment can be estimated from a relatively stable job: calculate parts produced per hour, multiply by the unit price, and subtract costs such as electricity and maintenance.
General-purpose robots complicate that calculation because their jobs can change. A robot might tighten screws on an assembly line one day and handle logistics or assembly work the next.
Agibot’s proposed answer is “Worken,” or effective intelligent work hours. It draws an analogy with tokens, the units used to measure content generated by digital AI models. Worken would measure useful work performed in the physical world.
Agibot calculates Worken by multiplying the number of active robots by their autonomous working hours and output efficiency ratio.
Active robots are those online and able to work during the period being measured. Autonomous working hours count continuous work without human supervision.
Agibot sets the output efficiency benchmark at 1.0: the acceptable output a skilled worker produces in a unit of time using a standard method. A robot’s actual output is divided by that benchmark. A ratio of 1.0 meets the standard, and a higher ratio exceeds it.
The proposed framework aims to make robot productivity comparable across industries and tasks.
Worken also points to a different way of valuing a robot. Its hardware depreciates, but models can be upgraded and skills expanded. Robots could also learn collectively, making their intelligence more useful even as their physical components lose value.
The purchasing decision would then hinge on how many hours of effective intelligent work each unit of spending generates over the robot’s lifetime.
Could customers pay for work instead of machines?
The announcements point to three changes in Agibot’s strategy.
The first is a move from individual technical advances toward a complete system: models for task execution, movement, and interaction, connected through AHA3 hardware and Link-U OS software.
“Developing the full stack in-house is not optional. It is a requirement,” Peng told a group of media outlets after the event. “The robot’s body sets the floor, and the model sets the ceiling. Agibot must have models that generalize well enough, bodies that are reliable enough, and an integrated hardware and software system that meets the standard. Deployment requires embodied intelligence companies to work across the full stack.”
He described Agibot as a company building the full stack around practical applications. “Our thinking is simple: make the technology reusable and make the products competitive enough. That is our core value.”
The second change is a move from selling robot hardware toward developing an industry ecosystem.
Agibot cited several deployments. At Longcheer’s Nanchang factory, eight G2 robots were integrated into a tablet quality inspection line and handled more than 17,000 products over six days, with a reported success rate of 99.99%.
At Chimelong in Hengqin, more than 500 robots were reported to be working routinely across seven types of application. At ASD stores, 100 X2 robots served as sales assistants and, according to 36Kr, helped generate RMB 2.23 million in sales over three days.
No single robot company is likely to serve every industry on its own. Agibot’s next challenge is to turn these early deployments into a wider range of repeatable applications.
Agibot is opening its operating system, development tools, and models through an ecosystem development program. The company plans to invest more than RMB 2 billion over five years, with RMB 100 million committed for implementation in 2026.
The aim is to give developers, universities, and commercial partners a shared technical foundation on which to build additional applications.
The third change is an attempt to influence how the industry measures value.
If customers use Worken in procurement decisions, they could begin comparing the productive work robots deliver, rather than simply the number of machines purchased.
Chuang Wang said in a post-event media interview that wider adoption could eventually shift pricing from buying equipment to buying productivity packages.
A customer might purchase a Worken allowance and schedule tasks. The provider could charge for actual use and combine several robots to serve different settings. Wang described this as a possible gradual evolution, with the business model still to be developed.
Agibot’s broader ambition is to turn laboratory demonstrations into infrastructure that produces measurable work.
“Every hot technology sector has its own cycle,” said co-founder and co-president Yao Maoqing, who also leads the embodied intelligence business. “For robotics, I think the companies that can go the distance will need the ability to turn technology and products into measurable productivity across the entire chain, and deliver quantifiable value in production.”
Whether customers will buy robot work this way depends on the test running through Agibot’s announcements: can the machines perform useful tasks reliably enough to justify the cost?
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.