COOWA Hires 'Real-Machine' Scientist to Monetize World Models with 50PB Data

Completing a few hundred sorting tasks at a trade show is just a demo. In the real world, a task must be executed thousands of times reliably, with customers paying for results, to qualify as a business.

COOWA Tech is applying this logic to push embodied intelligence into the deep end of large-scale deployment.

COOWA Tech is one of the few embodied intelligence companies in commercial operation, covering 50+ cities globally with over 10,000 robots in routine operation, and accumulating around 50PB of physical world data. This data is structured in three layers: dynamic scene data for world understanding, physical interaction data recording actions and feedback, and real-machine correction data for extreme long-tail scenarios.

However, data scale does not equal model capability. COOWA believes that the synergy of algorithms, hardware, and scenarios determines deployment ability, and the continuous acquisition and efficient conversion of real data directly impacts model iteration speed.

COOWA's self-developed COOWAM (Co-Optimized World Action Model) is the industry's first dual-layer agent world model architecture that natively integrates 'fast and slow thinking.' The upper 'slow thinking' Urban VLM handles long-term task understanding and planning, embedding a PVM (Productivity Value Measurement) module to judge whether a strategy creates commercial value before decisions—'whether what the robot does is worth doing.' The lower 'fast thinking' CooWAIM handles physical interaction inference and real-time correction within a 0–4 second window after actions—'how to do it specifically.'

The new chief scientist, Dr. Yang Xue, is a tenure-track assistant professor at Shanghai Jiao Tong University with over 13,000 Google Scholar citations and consecutively listed in Stanford's top 2% of scientists worldwide. Yang stated that he took the role to make his research impact the real world, not just citation counts. He will drive systematic evolution of COOWAM: empowering Urban VLM in multimodal and reward design, supporting CooWAIM upgrades in complex object detection and perception, and promoting cost reduction in the physical AI data factory through weakly supervised learning.

Future COOWAM will evolve from UMI and real-robot teleoperation data to first-person ego data and high-DOF dexterous manipulation. It will compress compute on edge via implicit representation distillation, compatible with wheeled robots, autonomous shuttles, quadruped robots, and humanoid robots.

COOWA's business loop is: acquire commercial data → low-cost data engineering extraction → leverage industry foundation models → new models deployed to feed back into business. Commercial use is not just deployment but also provides data sources, task goals, and value validation. Customers paying for results is the true test. The scaling law in the physical world cannot be defined in the lab; only real robots working and customers renewing can validate the value of this 'embodied brain.' Yang's arrival is key to making this faster, cheaper, and more sustainable.