Zhiyuan Links Two Foundation Models to a New Embodied AI Productivity Metric

Introduction

Embodied AI is moving beyond the question of whether a robot can complete a single demonstration. The harder question is whether it can work reliably, for long periods, with limited human intervention. At its Innovation Day, Zhiyuan Robotics announced the GO-3 Preview embodied foundation model, the WITA-OMNI multimodal interaction model, the AGILE motion-intelligence framework and Lingqu OS 1.0. It also proposed Worken, a physical-AI metric for describing effective intelligent work time.

Key takeaways

• Two models, two jobs. GO-3 Preview is aimed at task understanding, action generation and complex manipulation. It uses Zhiyuan’s embodied-native WRAM architecture and is supported by the Genie Evolver post-training system and Genie Master task framework. WITA-OMNI jointly models speech, vision and text inputs with speech, facial expression and motion outputs, targeting natural interaction in open environments.
• Reported benchmark performance. According to Zhiyuan, GO-3 Preview ranked first on RoboDojo, RoboTwin and RoboColiseum. WITA-OMNI reportedly topped Daily-Omni, TimeLens-Bench, EmbSpatialBench and MMAU. These results indicate strong performance on the cited evaluations, but they do not remove the challenges of changing environments, recovery from failures and continuous operation in the field.
• A tighter software-hardware stack. AHA3 is designed to coordinate motion, interaction and task intelligence through a single SoC. The G2 Max heavy-duty robot targets high-load, extended operation. Lingqu OS provides hardware abstraction, real-time control and reusable Skills. Zhiyuan says it plans to open-source Lingqu OS 1.0 and release parts of its production-validated code and software packages.
• Worken as a bridge to output. The proposed metric combines the number of active robots, autonomous working time and output efficiency. Its purpose is to put deployment scale, independence from human takeover and actual productivity into one frame.

Why it matters

The broader strategy is not simply to release another model or robot. Zhiyuan is trying to fill the engineering layer between research and commercialization: models interpret tasks and make decisions, robot bodies execute them, the operating system coordinates hardware and software, and field experience feeds back into training and planning. Deployments in manufacturing, retail and tourism are presented as practical tests of that stack.

Worken could also change the language of competition. Humanoid robots have often been judged by impressive motions, specifications or one-off success rates. If effective work time, uptime and qualified output are tracked consistently, the market may gradually focus less on spectacular demos and more on dependable productivity. Yet Worken is still a company-proposed exploratory metric. Comparable definitions across industries and independent validation of the data remain open questions.

Zhiyuan is therefore betting on integrated engineering across models, bodies and operating systems. The decisive test will not be the number of features announced, but whether robots can stay on the job, require fewer interventions and deliver acceptable results repeatedly.

Source: QbitAI