AGIBOT 发布量化机器人有用输出的工作框架

中文版正在随流水线自动整理，稍后刷新即可；可先阅读下方英文原文。

本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。

Robotics & Physical AI

AGIBOT Unveils Worken Framework to Quantify Robots’ Useful Output

Add Unite.AI to your preferred sources on Google

AGIBOT introduced Worken, a proposed framework for measuring how much productive work robots perform once deployed, at AGIBOT Innovation Day 2026 in Shanghai on October 29, 2026. The company also used the event to unveil its AHA3 integrated hardware platform, debut the GO-3 Preview embodied foundation model and the WITA-Omni interaction model, and formally launch the G2 Max industrial robot and Link-U OS 1.0.

The Worken Framework

The name combines the words “Work” and “Token,” and AGIBOT describes the concept as “Effective Intelligent Work Hours.” Under the proposal, Worken equals the number of active robots multiplied by autonomous working hours and by an output efficiency ratio.

The first factor counts robots that are actively performing tasks in real-world environments, not units that have merely been manufactured or sit in inventory. The second measures hours spent working autonomously rather than total powered-on time. The third scores output against a reference benchmark for a defined task, such as the amount of acceptable work a human worker completes under established operating standards.

Peng Zhihui, co-founder, president and CTO of AGIBOT, presented the framework as an answer to a measurement question he said the industry must address as robots begin working in factories, warehouses and service environments. “We want 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,” he said in the company’s announcement.

AGIBOT said it views Worken as an evolving framework that will be refined through real-world deployments and operational data, and that it hopes the measure can eventually become a broadly applicable reference for measuring the productive value of physical AI. The company also presents the measure as a way to examine the economics of robotics: while hardware depreciates over time, improvements in software and AI models may let the same machine perform more useful work during its operating life, so weighing output against total lifecycle costs could help businesses evaluate the longer-term value of a deployment.

Also released at the event was the “2026 Embodied Intelligence Industry Innovation Report,” a study of the industry’s progress toward real-world deployment jointly compiled by China Economic Information Service and the Shanghai Industrial Innovation Center of the China Academy of Information and Communications Technology. Worken and AGIBOT’s deployment examples appear in the report’s industry analysis.

AHA3 Hardware and New Physical AI Models

AHA3, short for AGIBOT Humanoid Architecture, is the company’s third-generation humanoid hardware architecture and, in AGIBOT’s description, its first commercial “Three Intelligences in One” platform, combining the hardware foundations for locomotion, interaction and task execution intelligence in a single architecture. The platform builds on advances in motor-drive technology, integrated computing and Wi-Fi 6 and 5G connectivity.

Alongside it, the company introduced a liquid-cooling system for the robot joints that it said improves thermal performance during continuous operation, specialized joints with IP67-rated protection for more demanding environments, a dexterous-hand family spanning the OmniHand 3 Lite, OmniHand 3 and OmniHand 3 Ultra, and the OmniPicker 3 tactile gripper.

In locomotion, the company detailed updates to AGILE, its multimodal, end-to-end motion intelligence framework, headlined by a new self-evolving agent capability. AGIBOT describes the addition as a shift from motion control toward spatial mobility intelligence, with the goal of helping robots adapt to the layouts and changing conditions of real deployment sites.

GO-3 Preview, presented by Dr. Yao Maoqing, co-founder, executive director, co-president and president of the embodied AI business, is a next-generation embodied foundation model built around WRAM (World Reasoning Action Model), a unified framework for reasoning, prediction and action. The company demonstrated zero-shot and few-shot task capabilities and described its Physical Intelligence Scaling approach, which pairs pre-training and post-training with continuous improvement after deployment. According to evaluation results the company presented at the event, the model took first place in both task success rate and process score on RoboDojo, achieved the highest overall score on RoboTwin and ranked first in overall score on RoboColiseum.

WITA-Omni, presented by Xiong Yan, partner, executive director, co-president, president of Central R&D and president of the interaction business, is an omni-modal model that brings visual perception, speech, real-time dialogue, facial expressions and physical actions into a single continuous interaction system. It supports full-duplex conversation with coordinated speech and movement and can be deployed on-device; AGIBOT said it currently ranks first on multiple embodied-interaction benchmarks, including Daily-Omni. The company said it is also expanding its data collection, simulation and training capacity through initiatives including AGIBOT WORLD 2026, the MEgo product family and Genie Sim 3.0.

G2 Max and Real-World Deployments

G2 Max is a heavy-duty industrial robot designed for manufacturing and logistics, with 21 degrees of freedom, a dual-arm handling capacity of up to 40 kilograms and an adjustable working height ranging from 0.1 to 2.5 meters. AGIBOT said the robot’s force-controlled manipulation, omnidirectional mobility and multiple safety features are intended to support efficient operations alongside existing industrial equipment and workflows.

The company pointed to G2-family deployments in tablet production at Longcheer, semiconductor packaging and testing, CNC machine tending and logistics operations. At Longcheer, robots handle material on the production line, with operating performance assessed against cycle time, sustained operation and successful task completion. Beyond manufacturing, AGIBOT highlighted Guangzhou Metro and the Chimelong tourism and entertainment complex, where robots perform tasks such as guiding visitors, answering questions and providing interactive services.

Chuang Wang, partner, executive director, senior vice president and president of the general-purpose humanoid business, outlined four capabilities behind scalable deployment: product readiness, manufacturing, on-site deployment and ongoing operations. AGIBOT said the contrast between factory and service work is why Worken relies on task-specific output benchmarks, and that it is using real deployment data and operating experience to explore how measures of autonomous working time and task output can be applied in each setting.

Link-U OS 1.0 and Open-Source Plans

AGIBOT announced the official release of Link-U OS 1.0, which it positions as the industry’s first operating system purpose-built for embodied intelligence. The system is built on a unified hardware abstraction layer, an embodied intelligence-native kernel and an intelligent service framework, connects robot hardware with upper-layer intelligent applications, and provides a full-stack foundation from development through deployment. According to the company, it has already been validated on mass-produced robots.

AGIBOT also announced an OpenAtom Foundation Platinum Level Sponsorship and the donation of the Link-U OS project to the foundation, with the two organizations planning to build an open-source community for humanoid robots, promote unified technical interfaces and standards, and collaborate on development, testing, validation and talent training. AGIBOT said its ecosystem now spans 255 universities, 511 enterprises and more than 80,000 developers. Link-U OS 1.0 is currently being trialled by selected ecosystem partners, with plans for it to become fully open source by the end of 2026.

Auto-fetched and translated by AI Atlas. All rights belong to the original author.