Dual-Layer Intelligent Agent World Model: Unifying Physical and Social Cognition

Introduction: The World Model Robots Truly Need

The world model that robots truly need is not a single physical world model, but the unification of a physical world model and a human social world model.

On July 17, the 2026 World Artificial Intelligence Conference (WAIC 2026) grandly opened at the Shanghai World Expo Exhibition & Convention Center. COOWA Technology made a stunning appearance at the Shanghai Guotou exhibition area. At the booth, the robotic arm driven by COOWAM—the full-stack self-developed dual-layer intelligent agent world model—smoothly completed long-range complex operations: the "Summer Special" beverage making demonstration. Meanwhile, the first heavy-duty quadruped robot dog X0 developed for complex urban scenarios also made its offline debut.

At the concurrent "Decoding the World" Development and Innovation Sub-Forum, Dr. Liao Wenlong, Founder and CTO of COOWA Technology, delivered a keynote speech titled "World Model: The Technical Foundation for Physical World Agents." At the conference, he disclosed to the industry the first dual-layer intelligent agent world model architecture—COOWAM (Co-Optimized World Action Model)—and based on this, outlined the ultimate vision of "RoboCity" driven by a unified model.

From technical demonstrations and new product debuts to revealing the underlying model architecture, COOWA Technology, deeply rooted in urban scenarios, is using world models as the foundation to promote urban embodied intelligence from individual operations to multi-agent collaboration, accelerating intelligence as a new productive force.

Opening Bottles, Cutting Watermelons, Making Special Drinks—The Battlefield of COOWAM World Model

At the booth, the robotic arm driven by the COOWAM world model continuously "worked" around the clock, making "Summer Special" drinks for the audience—cutting watermelon into pieces for juicing, unscrewing bottle caps, then adding lemon slices in a certain proportion, and pouring soda water. The entire set of movements was completely autonomous and smooth.

This task involves multi-robotic arm collaboration and the handling of physical laws such as contact forces and fluids. Under conditions where environmental factors such as lighting and cup position change dynamically, the robotic arm needs to perform real-time prediction and correction. Unlike traditional robots that rely on preset programs to repeat specific actions in fixed environments, robots based on COOWAM can deduce, make decisions, and act based on real-time environmental feedback while understanding task objectives, completing complex tasks with non-fixed processes.

This demonstration verified the technical feasibility of COOWAM as a universal foundation, supporting different robot forms to adapt to various physical interaction tasks, providing technical support for the implementation of embodied intelligent devices in scenarios such as urban public services and logistics distribution.

Quadruped Operation Platform X0 Debuts—Building the "All-Around Benchmark" for Urban Scenario Services

On the other side of the booth, the X0 robot dog, built based on COOWA's "Aether" heavy-duty quadruped platform, made its offline debut.

With a heavy load capacity of up to 68kg and all-terrain adaptive walking capability, X0 can easily break through the planar operation limitations of traditional wheeled equipment, extending the operation range to three-dimensional complex spaces such as stairs, slopes, gravel roads, and flower bed surroundings.

For X0, being able to walk and carry heavy loads is just the foundation. The value of "all-around" lies in penetrating into the capillaries of the city and achieving "one machine for multiple uses." With a highly modular design, X0 can quickly replace upper modules and seamlessly adapt to diverse tasks such as municipal sanitation, last-mile delivery, and security patrol. Currently, X0 has completed multi-scenario verification and entered the real machine deployment stage, applied to urban fine-grained operation and maintenance scenarios.

Reconstructing the Technical Foundation of "Urban Embodied Intelligence"—Disclosing the Industry's First Dual-Layer Intelligent Agent World Model Architecture

On July 18, Dr. Liao Wenlong, Founder and CTO of COOWA Technology, explained the design logic of the self-developed dual-layer intelligent agent world model COOWAM (Co-Optimized World Action Model) architecture in his forum speech. He pointed out that Physical Agents face two simultaneously existing and mutually coupled worlds: the natural world following physical laws and the human world following social rules.

Dr. Liao believes that for robots to realize application value, they need at least three capabilities: understanding natural laws (such as gravity, friction, inertia); deducing the results brought by different action choices before acting; and recognizing human social rules and industry SOPs (Standard Operating Procedures).

"The world model that robots truly need is not a single physical world model, but the unification of a physical world model and a human social world model," Dr. Liao stated. "Only by understanding both worlds simultaneously can Physical Agents truly step out of the laboratory, enter real society, and continuously create productivity value."

Based on this, COOWA developed the COOWAM dual-layer intelligent agent world model:

CooWAIM (Interactive World Action Model): Responsible for handling physical laws. This model focuses on the 0~4 second short-term future window, synchronously deducing environmental changes and correcting current actions when the robot generates actions, forming a decision-making closed loop of "action-future deduction-action correction."

Urban VLM (Human Social World Model): Built based on urban service data, responsible for understanding task semantics, industry SOPs, and productivity value orientation.

Inside Urban VLM, COOWA built PVM (Productivity Value Module), used to evaluate whether physical strategies meet task objectives, enabling robots to achieve continuous autonomous decision-making in understanding, reasoning, prediction, execution, and feedback.

Beyond the model architecture, Dr. Liao also mentioned a key question: How do world models become stronger? COOWA's answer is: through continuous interaction with the real world, letting the model continuously self-evolve in the data closed loop. Each task execution by the robot is a verification of the world model. What it sees, what actions it takes, how the environment changes, and whether the results meet objectives—these feedbacks will continuously feed the model.

Relying on normalized operations in more than 50 cities and regions worldwide, COOWA has currently established a commercial data flywheel of "real-world interaction-data accumulation-model learning-capability improvement," continuously optimizing world model performance through large-scale real machine deployment.

From Individual Intelligence to Interconnected Collaboration—Toward the Future City of "Robocity"

In 2026, embodied intelligence is crossing a turning point, gradually moving from purely technical exploration to large-scale application.

In COOWA's strategic blueprint, the maturity of a single Physical Agent or commercial scenario is just the beginning. When we truly achieve "one brain, multiple forms," letting the same COOWAM unified foundation drive various forms of wheeled robots, unmanned minibuses, robot dogs, and humanoid robots, our goal is no longer just to build a good automated device, but to reconstruct the operation mode of future cities.

This is COOWA's vision for the future—a "RoboCity" driven by a unified model.

In the future RoboCity vision, robots will bid farewell to the isolated state of performing fixed tasks, evolving into a network of Physical Agents (embodied intelligent agents) with autonomous decision-making capabilities and continuous value creation. Based on the shared "physical-social" unified world model, they will break information islands and build a truly tacit urban-level collaborative ecosystem.

To realize this vision, COOWA has planned a clear evolution path of "from outside the walls to inside the walls":

First Stage (Individual Intelligence): Deepen vertical fields. In scenarios such as urban travel, municipal sanitation, and property services, take the lead in completing the intelligent implementation of single devices.

Second Stage (Multi-Machine Collaboration): Expand composite scenarios. Driven by the unified world model, achieve efficient collaborative operations of intelligent agents in different forms in scenarios such as parks, communities, and industries.

Third Stage (Comprehensive Generalization): Move toward home scenarios. Overcome the barriers of complex operations and high generalization capability, letting embodied intelligence truly enter thousands of households.

This is also the ultimate goal of COOWA's continuous exploration of World Models: letting robots truly enter real society, weaving intelligence into every capillary of the city, and making Physical Agents a new type of underlying productive force driving the future RoboCity.

Source: QbitAI · July 22, 2026