Zhiyuan Unveils GO-3 Preview, WITA-OMNI, G2 Max Robot, AHA3 Architecture, and Lingqu OS 1.0 at Innovation Day

Hengyu sends from Aofensi
Qubit | Official WeChat Account QbitAI

Also fresh!

AGIBOT has come up with a way to give performance reviews to working robots (not really).

At this year's Innovation Day, AGIBOT co-founder, president, and CTO Peng Zhihui (Zhihui Jun) released a Physical AI new unit “Worken”, used to measure the effective intelligent working hours of robots.

AGIBOT also designed a performance settlement formula for Worken:

Worken = Number of active robots × Autonomous working duration × Output efficiency ratio

The number of active robots reflects the deployment scale, autonomous working duration indicates whether the robot can continuously operate without human intervention, and the output efficiency ratio measures the actual productivity efficiency.

Fine, from now on silicon-based people can no longer just slack off ╮(╯▽╰)╭

However, this time AGIBOT is not only arranging a performance evaluation for robots.

Focusing on how to enable robots to truly create value, AGIBOT has also brought a series of technical advances ranging from models and bodies to operating systems, covering core capabilities such as motion, operation, and interaction in embodied intelligence.

For example, on the new model side, the embodied base model GO-3 Preview has already secured first place in three embodied intelligence evaluation rankings, and the interaction large model WITA-OMNI has also topped four evaluation rankings.

After the meeting interview, Zhihuijun stated “Ten thousand units offline, is just a victory for manufacturing; ten thousand units on duty, is the victory of physical AI”.

But to let robots continuously deliver this performance, just having a set of calculation formulas is not enough.

AGIBOT's new model, new ontology, and new system released this time can exactly connect the technological layout behind Worken.

From model to ontology, what new things has AGIBOT released?

At this innovation day, AGIBOT released the new generation of embodied base model GO-3 Preview and the full modality end-to-end interaction large model WITA-OMNI.

The two models respectively target the two core capabilities of robot work.

GO-3 is mainly oriented toward operational intelligence, helping robots understand tasks, generate actions, and complete operations; WITA-OMNI is oriented toward interaction intelligence, enabling robots to understand people, respond to them, and engage in natural communication through voice, expressions, and gestures in real-world environments.

Meanwhile, AGIBOT is also advancing the upgrade of motion intelligence, hardware architecture, and operating system.

Models are responsible for enhancing capabilities, ontologies and systems are responsible for carrying and coordinating capabilities, and real deployment tests whether these technologies can play a role on the job site.

Thus, AGIBOT has established a “Triple Intelligence” layout encompassing interactive intelligence, motion intelligence, and operational intelligence.

Job Intelligence

First look Job Intelligence AGIBOT's latest contestant is the embodied base model GO-3 Preview.

It has taken the top position in three embodied intelligence evaluation rankings: RoboDojo, RoboTwin, and RoboColiseum, achieving a triple top ranking.

It is reported that GO-3 Preview adopts the embodied native architecture WRAM（World Reasoning Action Model）combining task understanding, reasoning, future state prediction, and action generation.

When robots perform fine operations, they need to understand the task goal, plan actions, and also predict the changes that actions may bring. WRAM attempts to integrate these processes into a unified learning framework, enhancing the model's ability to handle complex tasks.

In the pre-training phase, based on joint scaling of models and data, GO-3 Preview exhibits stronger task transfer and complex skill adaptation capabilities, thus performing better in Zero-Shot and Few-Shot tasks.

No robot can escape the test of LEGO!

After fine-tuning, robots equipped with GO-3 Preview can complete tasks such as part grasping, alignment, and insertion, and can complete assembly tasks under the coordination of an agent.

In addition, the model also supports cross-body transfer. By switching to the bipedal humanoid robot X2, it can also open doors with its hands and kneel to pick up objects.

But as you know, the laboratory environment is relatively simple. In reality, the positions of objects and environments change endlessly, and embodied models cannot rely solely on pre-training to learn all their skills.

AGIBOT has therefore constructed a post-training system Genie Evolver, combining 100 real machines, physical simulation, and world models, which is equivalent to giving the robot an internship workshop.

When entering actual positions, the embodied intelligence framework Genie Master continues to take charge of task planning and anomaly recovery, turning today's mistakes into tomorrow's experience.

General capabilities rely on training - improving through virtual and real interaction - reusing real experience, forming a chain of processes, and embodied robots are completely different from traditional equipment.

Interactive Intelligence

If GO-3 focuses on how robots complete work tasks, WITA-OMNI focuses on how robots interact with humans.

As AGIBOT's newly released interactive large model, WITA-OMNI has won the top position in four authoritative evaluations: Daily-Omni, TimeLens-Bench, EmbSpatialBench, and MMAU.

WITA-OMNI models the input speech, vision, and text, along with the output speech, expressions, and actions, within the same model, enabling the robot to listen, watch, speak, and move simultaneously.

For example, in open human environments such as hotels and tourist attractions, robots need to determine who is interacting with them, whether the other party has finished speaking, and when it is appropriate to respond, in noisy environments.

One interesting feature is that WITA-OMNI does not require a wake-up word.

AGIBOT enables the model to identify the conversation partner by integrating information such as sound direction, voiceprint, face, and lip movement, and then determine the timing of interaction through gaze and tone. Even if the robot is performing other actions, it can continue to participate in the conversation.

Okay, back to the main topic!

WITA-OMNI adopts a fast-slow dual system design.

• The front-end Interaction Engine is responsible for immediate responses.
• The back-end Background Agent handles reasoning, planning, tool invocation, and long-duration tasks.

When encountering complex problems, the backend continues to process while the frontend maintains the conversation and synchronizes progress, allowing users to supplement requirements or adjust tasks at any time.

During the live demonstration, AGIBOT partner, co-CEO, and head of the Interactive Business System Xiong Yan fired rapid-fire questions, and the robot equipped with WITA-OMNI answered fluently; when asked “Who is the cutest in AGIBOT,” it took the opportunity to indulge in a bit of self-love.

According to data disclosed by AGIBOT, WITA-OMNI typically responds within an average of 600 milliseconds, with the fastest response around 400 milliseconds. This speed is already close to natural reaction times in human conversation.

In terms of language, WITA-OMNI supports understanding over 100 languages and various dialects, can speak 29 languages and 7 dialects, and supports local offline interaction without internet access.

Another interesting feature is Intelligent FaceUI.

The user just needs to ask a question, such as 'Tell me about the Pythagorean theorem,' and the robot can dynamically display information on its facial screen.

When encountering complex issues that cannot be explained clearly at once, graphical representation is obviously more intuitive than dry voice announcements.

AGIBOT plans to complete full domestic deployment by December 2026, launch the overseas version in Europe in January 2027, and complete full overseas regional rollout by March.

Motion Intelligence

In addition to job intelligence and interaction intelligence, AGIBOT is also continuously advancing innovation in the field of motion intelligence.

The AGILE general-purpose motion intelligence framework released at Innovation Day incorporates environmental perception, terrain understanding, whole-body motion control, and upper-lower limb coordinated operation into a unified perception and control loop.

Combined with the self-evolution architecture, enabling motion intelligence, as a foundational intelligence, to evolve from motion control to space-moving intelligence capable of autonomous evolution, better meeting the spatial movement requirements in deployment scenarios.

When facing dynamic obstacles, multi-person collaboration, and unexpected human-robot interactions, robots need to adjust their actions according to environmental changes while ensuring safety.

New architecture and new body

In the early stages of the industry, most robots had only a single intelligent capability. Later, many solutions moved towards separating the three intelligences, i.e., “three sets of intelligence running on their own independent hardware boards,” which is also the route many vendors are currently trying to pursue.

AGIBOT has taken another step forward by introducing the third-generation hardware architecture AHA3Integrate the three sets of intelligent functions—movement, interaction, and task execution—into a single control system.

AHA3 is uniformly scheduled by a single SoC, achieving the integration of the three intelligences.

According to the data disclosed by AGIBOT, this upgrade reduces the physical size by 50% and improves the response speed by 90%.

According to the algorithm of the performance evaluation formula, all the intelligence of the robot can only be realized as Worken through a reliable body.

Dei, AGIBOT also released a new robot body: the safety-rated heavy-duty embodied robot, the G2 Max.

It supports a single arm load of 20 kg, dual arm peak load of 50 kg, equipped with seven-degree-of-freedom force-controlled arms, and through multiple safety protections and dual battery hot-swapping designs, aimed at long-term continuous operation needs.

Also released at the same time is the world's first subcutaneous joint liquid cooling system and the new generation of dexterous hand family, further solidifying the foundation of the body and core components.

Using an embodied native OS to integrate robot software and hardware

As models and robot bodies continue to upgrade, how to make different hardware, underlying control, and upper-level intelligent applications work efficiently together has become another engineering challenge.

AGIBOT's solution is to show off its embodied native OS.

Ding ding ding ding!

AGIBOT released the embodied intelligent native operating system, Lingqu OS 1.0～

This embodied intelligent native operating system covers the entire process from construction, deployment to operation. It is responsible for hardware access and operation at the lower level, and provides an application runtime environment for motion intelligence, operation intelligence, and interaction intelligence at the upper level. It has already been verified by mass-produced robots.

It can't be said not to be pressure mountainous...

AGIBOT stated that Lingqu OS first solves the problem of hardware access.

It supports the rapid connection of 100+ peripheral devices across 15 categories through a unified hardware abstraction framework, and is compatible with mainstream computing platforms.

For hardware developers, new devices can be connected to the system faster, reducing the workload of repeated adaptation, and leaving space for innovation for different hardware manufacturers.

Another key point is real-time performance.

When robots perform high-frequency motion control tasks, the system must not only respond quickly, but also ensure that the response time is sufficiently stable. That is, the system needs to minimize the time fluctuation each time it executes a command, allowing the robot to complete the action as expected.

Lingqu OS has performed determinism delay optimization for the entire stack, with end-to-end delay jitter less than 4 microseconds under 1000Hz high-frequency task load.

On this basis, Lingqu OS also provides an intelligent service framework, supporting mainstream inference frameworks, and encapsulating system services as reusable Skills.

Developers can directly call pre-packaged system capabilities (Skills) to build applications without having to start from low-level adaptation for every task.

Moreover, AGIBOT has arranged this thing on the open source server!

By the end of this year, AGIBOT will fully open source Lingqu OS 1.0 and donate the project to the OpenAtom Open Source Foundation.

According to the plan disclosed at the main forum, AGIBOT will also open up 200 million lines of code and 5000+ ecosystem software packages that have been verified through mass production, provide an end-to-end construction toolchain, and pre-integrate various open-source components for embodied intelligence.

Those who do not use AGIBOT's own system products can wait to experience how good Lingqu is with their own software and hardware.

I'll sit on the floor and wait for your feedback~

(chewing.gif)

Paper strength converted into KPI, robot performance review

No matter how complete the technical route is, the robots still need to work continuously on the job site in the end.

Now, AGIBOT has deployed robots in industrial, retail, and cultural tourism scenarios, and the requirements for them vary across different scenarios.

In Industrial scenario where the most challenging aspects are task accuracy, operation rhythm, and running stability.

Currently, AGIBOT and Longcheer (a leading global consumer electronics ODM manufacturer) have collaborated to build the world's first complete 3C production line covered by embodied intelligence.

This production line achieved large-scale production in March 2026, and as of the time of the innovation day conference, it has been running stably for more than 7 months.

During a 6-day full production line live stream previously, the robots completed over 64,824 operations, producing 17,625 tablet computers, with an operation success rate of 99.99%.

The average operation time per item is 22 seconds, comparable to that of a human skilled worker.

Looking at the retail scenario, which is the most challenging in terms of interaction understanding and task changes Retail scenario.

In the offline stores of Aishida, 100 Lingxi X2 robots acted as silicon-based guides, generating sales of 2,230,000 yuan in three days.

Cultural tourism scenario faces changing crowds, noisy environments, and multi-language communication, and must run continuously during daily operations.

Hengqin Changlong Spacecraft Park has deployed nearly 500 robots, covering over 100 positions, providing explanations and interactive services, with an average of about 10,000 visitors per day during the National Day holiday.

Now, at the Jingqi World exhibition area, robots can already generate real-time commentary visuals using WITA-OMNI, placing visitors directly on the front line of intelligent experiences (doge).

The 10-month Worken dashboard displayed at the main forum lists data from scenarios such as Longqi production line, Guangzhou Metro, Hengqin Changlong, and Aishida retail stores.

It's impressive, and embodied intelligence has been running wild like a loose horse for so long, whether in deployment scale, autonomous working duration, or output quality, there's finally a tentative quantitative approach being explored.

Worken is one such attempt.

Just as Token gradually became an important unit for measuring model usage and cost in the era of large language models, Worken aims to provide a commercial anchor for physical AI that is tied to actual work, shifting industry competition from who has the cooler demo and more impressive concept, to who has robots that can continuously and stably create productivity.

Therefore, the combination of AGIBOT's new model, new ontology, and native OS aims to revolve around Worken, transforming technical capabilities into effective working hours as much as possible.

R&D certainly has long-term value, not every innovation needs to turn into effective working hours immediately.

But embodied intelligence ultimately depends on whether it can continuously create value in real-world tasks. And Worken is that measuring stick.

Hardware depreciates, but intelligence can appreciate. The field of embodied intelligence is ready to move beyond the stage of a greenhouse performance and prove itself on the real battlefield of jobs.