Games and AI, evolving together
Where will AGI be born? Alaya Lab has put the answer in the game.
From AlphaGo's breakthroughs in the game of Go, OpenAI Five's defeat of the Dota 2 world championship, to Google DeepMind's expansion of research into large open world multiplayer online games such as EVE Online, and the recent heat of Vibe Game and World Models, games continue to provide AI with a more complex, open environment that brings challenges closer to or beyond the complexities of the real world.
At the same time, generative AI has begun to reshape game engineering in turn.
From planning assistance, automation testing, to 2D, 3D, music and other asset creation, to the rapidly developing programming intelligence, generative AI is accelerating integration into all aspects of game research and development, continually improving production efficiency and content quality.
Alaya Lab believes that gaming is not only an entertainment product, but also an important route to AGI.
The next generation of games will have a larger world, richer interactions, and a continuously evolving human-biotechnology ecosystem. It will be an important environment for AI to learn and evolve, and it will also be a whole new space for human co-creation and co-evolution with large-scale intelligent agents.
Based on this vision, Alaya Lab has officially released Project Gear, which explores the next generation of game engineering and game formats, making games a super-engineering drive for intelligent advancement.
Gear consists of two core components:
Gear Engine: Geared towards game production and exploring the shape of the next generation of game world. Beyond existing game AI applications such as programming, asset generation, automation testing, Alaya Lab is focused on building, operating and continuing to evolve the next generation of game worlds based on hybrid world models.
Gear Platform: A multi-person, multi-smart collaboration to explore how large-scale people and smart bodies work together on complex projects. Among them, Gear Zero is geared towards mass creators and Gear Cadre is geared towards professional teams.
Alaya Lab hopes to push the boundaries of the next generation of game engineering through Gear, connecting AI, games and creators.
The official website:
Gear Zero test experience address
Gear Engine: How to remake games
Alaya Lab focuses on building and running the next generation of gaming worlds and explores game production and operating paradigms through Gear Engine.
On the one hand, Gear Engine proposes World Compilation, which explores explicit world modeling, which automatically composes the visual world into a structured, physically-interacting, logical, and directly executable world system; on the other hand, it explores implicit world projections through video world models, achieving high degrees of freedom in real-time generation and retrieval, and focuses on solving key challenges such as controllability, image quality drift, time-space consistency, and real-time generation.
On this basis, Gear Engine is further advancing the integration of optical modeling and implicit projection, exploring the next generation of hybrid game world models: both maintaining the controllability, stability and executability of the optical world, and leveraging the advantages of generating models over creating large-scale content and generating an open world, gradually building an AI-driven game production and operating system.
World Compilation "Modeling the Opinions of the World"
World Compilation revolves around three levels of Structure, Interaction, and Program, and compiles the visual world into explicit, structured, editable, and executable expressions.
At the Structure level, WorldSculpt is able to reconstruct the apparent structure of the world from the video, restore the geometry, spatial positioning and interrelationship of the objects in the scenes. WorldSculpt's output is aligned with the video, without the need for additional optimization for the object layout. It is therefore more suitable for the game's large-scale, complex scene construction (see detailed technical report Benchmark). WorldSculpt is also able to break down three-dimensional marble scenes generated by World Labs' Marble models into independent objects, so that the original holistic three-dimensional representation can be transformed into a class of objects that can be operated separately, assigned attributes, and represented interacting.
On this basis, PartTrellis further analyzed the internal structure of the object, identifying and breaking down its moving parts, thus forming a layered representation of the structure from scene, object to component, laying the foundation for the subsequent physical attribution and interaction logic construction.
At the Interaction level, the Agent further builds interactive properties for the world, including joints, collisions and other physical parameters, and models the dynamic behavior of objects and roles in the interaction process. At the same time, through self-improving mechanisms, the Agent can continue to learn from the failure of execution, the simulation of results and environmental feedback, the continuous improvement of the modeling and generation of physical properties, binding relationships and interactions, and the reduction of the token expense of iteration.
At the Program level, the system further organizes the rules of structure, physics, and interaction into executable program expressions, generates and maintains simulation code with persistent state by Agent, and combines procedural generation to efficiently construct and extend large-scale outdoor environments, ultimately transforming static visual content into an executable, interactive, and sustainable evolving world.
This is a video of the "Invisible World Tour".
Recently emerging models of the video world provide another implicit modeling route for building the gaming world. It skips traditional processes such as asset production and code development to use video images as an expression, and models directly output user-expected images and interactive feedback.
In this direction, Alaya Lab launched Alaya-World alongside Alaya-EVOKE in July, aimed at addressing the critical challenges of video world models in controllable interaction, time-space consistency and gravitational drift to achieve continuous interaction at hourly stability and high frequency.
Among them, as of August 15, Alaya-EVOKE is at the top of the WBench World Model Rating list.
In July, Alaya Lab also introduced ShadowDancer, a simulated action model for gameplay interaction that expands the concept of "action libraries" in traditional games into world models, and continues to expand through efficient action extraction and action libraries, enabling world models to achieve rich interaction at the game level.
In addition, Alaya Lab has developed a simulation data engine to scale up the first-person/third-person/full-scene video from UE with corresponding real-valued depth, white mode, light flow, point tracking and other geometric information to support the training of video world models.
The "mixed world" model
While the world model has attractive prospects for high freedoms, real-time generation and visual performance, there are still significant limitations on long-term stability, logical follow-up, gameplay generalization and running costs that are difficult to fully break through in the short term. While optical modeling achieves content sustainability, logical consistency and operational efficiency through structured expression and code, it can complement video world models.
In April, Alaya Lab introduced the generated world-rendering Alaya R.