Zhiyuan Robotics Unveils WITA-OMNI Model, Lingqu OS 1.0, and AIMA Ecosystem at October 29 Innovation Day

Robotics Outlook (WeChat Official Account: robot_pro)
Author | Xu Lisi
Editor | Mo Ying

Are robots also starting to clock in for work?

Calculating working hours and setting KPIs for robots: how much is a robot's day of work actually worth? AGIBOT aims to set a standard answer to this question.

On October 29, the 2026 AGIBOT Innovation Day was held at the Shanghai Expo Center. After attending the launch event, the information density was truly quite high!

AGIBOT has proposed a new unit of measurement—Worken (Effective Intelligent Work Hours), which uses the number of active robots, autonomous working duration, and output efficiency ratio to measure exactly how much work robots have done.

The GO-3 Preview, released immediately after Embodied Base Model ranked first in all three evaluation lists: RoboDojo, RoboTwin, and RoboColiseum. Embodied Intelligence will be open-sourced by the end of the year., Subsequently,

Subsequently,, took the stage and unveiled multiple achievements in one breath:

Embodied Full-Modal End-to-End Interaction Large Model WITA-OMNI officially released, This model has topped four interaction-related evaluation leaderboards;

The industry's first native embodied intelligence operating system Lingqu OS 1.0was officially unveiled and donated to the OpenAtom Foundation. AGIBOT plans to fully open-source Lingqu OS 1.0 by year-end.

On the ecosystem side, the latest results of the AIMA technology ecosystem and the YuanSheng initiative were also announced simultaneously, further opening the technical foundation to the entire industry.

After the press conference, we had a conversation with Xiong Yan. Through this interview, we found that AGIBOT's series of launches all point to one thing: enabling robots to truly enter workplaces and transforming model capabilities into sustainable productive forces that create value.

I. Robots can now read the room! WITA-OMNI overcomes the challenges of real-world interaction.

When a robot stands in a crowd and hears someone speaking, should it respond? Not necessarily.

For example, in an amusement park, there are announcements, music, tourists chatting, and children who may interject at any moment. The robot must not only hear the sounds clearly but also judge whether someone is addressing it, when to respond, and even use gestures and expressions to reply.

Xiong Yan explained that natural interaction is difficult due to robustness, emotional intelligence, interaction experience, and intelligence. WITA-OMNI 1.0, AGIBOT prioritizes robustness, focusing on improving recognition in noisy environments and during multi-person conversations, while also enhancing understanding and response to emotions.

1. Clear listening, seamless engagement, and rapid response. Native full-modal architecture. Model debut.

WITA-OMNI follows a native full-modal technical route: It integrates inputs such as speech, vision, and text, along with outputs like speech, facial expressions, and gestures, into a single model. Through unified training and alignment, the robot can make judgments by combining visual and auditory information and coordinate verbal and physical actions accordingly.

Why design it this way?

Xiong Yan explained, 'Devices such as smartphones, computers, and smart speakers are typically interacted with via microphones. However, robots need to integrate directly into real-world crowds; beyond visually perceiving the surrounding environment, they must combine language, vision, and other information to continuously understand context throughout interactions.'

For example, when a robot hears someone speaking, it is difficult to determine solely from speech whether that person is addressing itself; visual information is also needed to identify the speaker and their interaction intent.

He revealed that the team attempted adding new modalities after an existing large model, but this often caused loss of original capabilities and made it hard for newly added information to align sufficiently with original training results. Therefore, AGIBOT chose to introduce various modalities during the initial training phase, enabling the model to better grasp conversational timing, understand intent, and judge when to interrupt or remain silent.

Interaction must also be sufficiently fast, so WITA-OMNI adopts full-duplex and dual-mode fast-slow Architecture. When discussing this design, Xiong Yan cited human communication as an example:"Human interaction systems are naturally duplex; I can listen while expressing myself."

The foreground of WITA-OMNI handles immediate responses, while the background manages complex reasoning, planning, and task processing; both share context and timelines. Thus, even while the background processes tasks, the robot can maintain communication, allowing users to supplement or modify instructions at any time.

On-site, we saw WITA-OMNI demonstrate its outstanding capabilities in human-machine interaction with the Expy A3. It is understood that AGIBOT equips the Expy A3 with multiple cameras, LiDAR, eight microphones, and a high-definition face screen, enabling the robot to communicate by combining vision, hearing, and body movements.

The Expy A3 has learned to read expressions; when Xiong Yan asked the robot what it saw, the Expy A3 replied that he was an "uncle" standing on stage. Xiong Yan then retorted: "Who do you say is your uncle?" The robot immediately changed its wording and called him a "handsome guy".

The Expy A3 can also withstand various rapid-fire follow-up questions, and simultaneous interpretation in multiple languages is no problem. According to Xiong Yan, the model's average response speed is about 600 milliseconds, with a fastest speed of 400 milliseconds.

Additionally, WITA-OMNI also in the four authoritative evaluation benchmarks Daily-Omni, TimeLens-Bench, EmbSpatialBench, and MMAU, ranking first on The evaluation leaderboard achieved first place..

2. Enter Chimelong Paradise to validate interaction capabilities in real-world scenarios.

In multiple real-world scenarios, the interaction experience brought by WITA-OMNI is unprecedented.

Taking Chimelong Paradise as an example, Xiong Yan introduced that some children were too shy to speak up, and standing robots could not hear them clearly; however, the Expy A3 would actively approach to greet them, then kneel on one knee and bend down to reduce the distance. 'I can actively approach the interaction target, which is a very different interaction process that ordinary models cannot complete.'

When robots actively approach people, safety becomes the top priority. Chimelong Paradise receives a large number of tourists daily, and children may suddenly run into the robot's activity area.

In response, the Expy A3 can maintain 360-degree environmental observation; actively stop movement when someone enters the safety zone; quickly reduce output force upon encountering accidental contact. Electronic fences restrict the activity route, and abnormal alerts notify operational personnel to intervene.

Robots at Chimelong Paradise have cumulatively completed more than 1,000,000 interactions., Large-scale real-world interactions also provide the model with opportunities for continued learning. "Real deployment data is the most valuable to us; this data itself constitutes a significant moat," said Xiong Yan.

Compared to smoothly completed conversations, data such as user dissatisfaction or mid-conversation departures that flow back into model training enable robots to gradually become familiar with actual service environments.

More importantly, improvements in interaction intelligence are directly linked to whether robots can create tangible economic value.

"Interaction may be the essential gateway through which humans generate value."Xiong Yan believes. For instance, in comprehensive scenarios such as hotel services, shopping guidance and tours, and retail warehouses, robots must first use interaction to understand what users want. Only after that do motion intelligence and operational intelligence take over.

II. Building an Operating System for Robots: Lingqu OS Will Be Fully Open-Source by Year-End

Even with a smarter brain, robots still face challenges in hardware-software adaptation.

"The embodied era requires a native operating system. For AI to move beyond the screen—to see, hear, act, and work—hardware and software must collaborate efficiently. Without such a unified system, our development would be repetitive, complex, and inefficient," said Xiong Yan.

The Lingqu OS 1.0 released by AGIBOT this time aims to address these issues. It achieves unified hardware abstraction, an embodied native kernel, and an intelligent service framework, covering bottom-layer hardware access, software construction, deployment, and upper-layer application support, providing a unified software foundation for different robots.

This system has already been validated on AGIBOT's mass-produced robots. Lingqu OS Supports mainstream computing power platforms, and enables rapid access to over 15 categories and 100 types of peripherals, reducing redundant adaptation work between different devices.

The robot operating system must also address real-time challenges; while the robot is in motion, the system needs to continuously process control commands, where latency fluctuations may affect action execution. To meet this requirement, Lingqu OS has implemented full-stack deterministic latency optimization,, achieving end-to-end latency jitter of less than 4 microseconds under a 1000Hz task load.

More importantly, Lingqu OS provides a unified access platform for three types of intelligence: motion, interaction, and operation. Developers can invoke and reuse existing system capabilities to develop robot applications tailored to different embodiments, reducing redundant construction.

Xiong Yan also announced on-site that by the end of 2026, Lingqu OS 1.0 will be fully open-sourced, contributing 200 million lines of production-verified code, over 5,000 ecosystem software packages, an end-to-end build toolchain, and pre-integrated open-source components.

Additionally, AGIBOT has donated the Lingqu OS project to the OpenAtom Foundation and will jointly build a humanoid robot open-source community. The collaboration will cover operating systems, embodiment hardware, data, models, and toolchains, promoting joint development, testing and validation, and industry standard construction.

Behind the open-source initiative, AGIBOT hopes for a more prosperous technical ecosystem in the industry.

III. In half a year, 80,000 developers have gathered; AGIBOT opens its technology foundation to the industry.

In April this year, AGIBOT launched the Yuan Sheng initiative, aiming to open up its own technologies, products, and engineering capabilities to promote the development of the AIMA ecosystem.

Behind this judgment lies the industry's real predicament: the bottlenecks in embodied intelligence have long gone beyond hardware, and also lack scenarios, lack data, and lack developers who can put robots to use.

By opening up its technical foundation, AGIBOT is essentially trading its own investment for the entire industry's speed of application trial-and-error; the larger the ecosystem, the higher the probability of successfully running through scenarios.

Xiong Yan revealed that the AIMA ecosystem currently has gathered over 80,000 developers, with open-source dataset downloads exceeding 5,600,000 times, Lingxin and Lingchuang platforms cumulatively generating 150,000 user-created content items, of which over 7,000 have been actively shared by creators on the platform.

In detail, the Lingxin platform has cumulatively generated over 30,000 agents, supporting developers in agent creation and natural language orchestration; the Lingchuang platform has accumulated over 30,000 high-dynamic actions, supporting minute-level replication of full-body robot motions and reducing the heavy programming work involved in the creation process.

Beyond application development, robots also need to collaborate. AGIBOT's Lingdong Platform can uniformly schedule different robot models and has already been used for the daily human-robot dance performances at Chimelong Paradise. On October 24, 110 Expy A3 units, leveraging this system, completed a Guinness World Record challenge for the largest number of humanoid robots dancing simultaneously.

In industrial operations, Genie Studio 2.0the platform has been upgraded to an intelligent operation platform, capable of deploying one type of operation scenario in as little as seven days. The CoLab initiative behind it has driven over 100 ecosystem co-building projects to implementation, covering more than 50 application scenarios, further applying platform capabilities to actual delivery.

Behind the ecosystem expansion lies substantial financial investment: since 2026, Yuan Sheng has invested over 300,000,000 yuan, with a five-year plan committing no less than 2,000,000,000 yuan. Currently, AGIBOT has connected with 255 universities and 511 enterprises; in the past half-year, 40 partners received deployment-stage implementation incentives.

Conclusion: Open Ecosystem, Promoting the Prosperity of Embodied AI Technology

"For an industry to develop and technology to grow, the ecosystem is the soil."This statement by Xiong Yan at the launch event also highlights the thinking behind AGIBOT's current series of strategic layouts. For robots to truly enter more job roles, they require not only continuous improvement in model capabilities but also the joint participation of developers, scenario partners, and the entire industrial ecosystem.

This time, AGIBOT plans to open-source Lingqu OS and continue investing in the AIMA technology ecosystem, aiming to reduce redundant development across the industry by opening up technology and engineering capabilities, enabling more innovative outcomes to be applied and validated.

An open ecosystem is expected to accelerate the sharing, reuse, and innovation of technological achievements, involve more developers in technical exploration, validate more application scenarios, and drive the industry toward a virtuous cycle of continuous evolution. As technological innovation and industrial applications further integrate, embodied AI will usher in a more prosperous technology ecosystem.