GigaAI’s General World Model at WAIC: From Generation to Action – The Path to Physical AGI

World Model Exhibition

The GigaAI booth at WAIC 2026 attracted continuous crowds, media, industry customers, and even overseas dignitaries, highlighting that the attention was not merely because the robots could move.

What drew the focus was the display of a complete General World Model product matrix: YiSu for content creation, DriveDreamer for autonomous‑driving data and simulation, GigaWorld for embodied intelligence, GigaBrain and GigaWorld‑Policy for turning world understanding into actions, Shiguang S1 for household scenarios, and Maker H01 for intelligent manufacturing.

Generation Models

YiSu combines a real‑time multimodal world model, an agentic model that understands physics, and an AI‑native creation engine, enabling users to script scenes, performances, and camera moves with natural language for film, game, and interactive content production.

DriveDreamer targets autonomous‑driving data and simulation. As real‑world road testing becomes increasingly costly, the world model provides a generative, verifiable, and iterative training and evaluation environment for intelligent driving systems.

GigaWorld serves embodied intelligence by supplying large‑scale training data, simulation environments, and closed‑loop verification. The latest GigaWorld‑1 pushes the model from merely understanding the world to testing actions within it.

Action Models

GigaBrain is a general embodied brain that handles cross‑entity, multi‑task, long‑horizon real‑world operations.

GigaWorld‑Policy attempts to output actions directly under the world‑model paradigm, giving robots physical foresight before execution and moving toward real‑robot deployment.

Together, GigaWorld generates and simulates the world, GigaBrain interprets tasks and generates strategies, GigaWorld‑Policy translates predictions into actions, and real‑world feedback loops back to the model.

Family Scenario

In a recreated home environment, Shiguang S1 equipped with GigaBrain performed continuous tasks such as fetching meals, clearing tables, making coffee, sorting clothes, and handling parcels, demonstrating the system’s ability to maintain stable execution amid long‑tail variations in lighting, occlusion, and user commands.

The key question was not whether a robot could demonstrate a single action, but whether an embodied brain could continuously understand, plan, and execute long‑duration operations in an open, dynamic household.

Smart Manufacturing

Maker H01 demonstrated autonomous loading, transport, and joint assembly tasks in an industrial setting, highlighting the high‑frequency, standardized nature of these tasks and their suitability for quantitative evaluation of efficiency, yield, and cost.

GigaAI’s concept of “robots building robots” envisions a realistic deployment path: robots first participate in manufacturing other robots, then gradually expand to more flexible production tasks, creating a feedback loop where physical AI continuously refines its models through real‑world operation.

Industry Insight

The exhibition emphasized that a world model is not merely a video‑generation tool but the foundational infrastructure for AI to enter the physical world. The competitive focus has shifted from visual quality to answering whether actions are correctly understood, states remain continuous, physical outcomes are credible, and strategies transfer to real environments.

Consequently, the barrier for general‑world‑model companies lies in the efficiency of the closed‑loop system comprising model, data, embodied hardware, evaluation, and scene deployment.

GigaAI’s product lineup exemplifies this: YiSu and DriveDreamer showcase generation capabilities; GigaWorld and GigaWorld‑Policy demonstrate the transition from generation to action; GigaBrain provides embodied decision‑making; Shiguang S1 and Maker H01 validate the models in household and industrial contexts.

The ultimate goal is to enable AI to understand not only text and images but also space, objects, actions, contact, failure, and feedback—moving physical AGI from generating realistic videos to reliably guiding agents in the real world.