Embodied World

• 

• 

Inspiration

We are developing systems that generate designs for electrical and mechanical products. These systems can already produce components such as bills of materials, wiring plans, system architectures, and visual concepts.

The next challenge was understanding how those designs would behave and appear in the physical world.

A generated machine should not exist only as a diagram or a collection of parts. We wanted to place it inside an environment where it could be observed at scale, tested spatially, and eventually interacted with by people, robots, and autonomous agents.

That idea became Embodied World: a platform for turning generated hardware concepts into explorable, persistent, and interactive environments.

What it does

Embodied World transforms a product concept, image, or generated hardware design into a spatial environment that users can explore.

The platform is designed to combine two complementary representations:

• An explicit 3D environment containing geometry, objects, collision data, and spatial relationships.
• A generative simulation layer capable of producing atmosphere, motion, environmental changes, and previously unseen regions.

A user can generate or import a machine, place it inside a virtual environment, move around it, inspect its scale, and understand how it might exist in a real operational setting.

In the future, these environments could also be used to test assembly processes, maintenance workflows, robot navigation, manufacturing layouts, and human interaction with generated systems.

Embodied World is intended to act as a bridge between generative design and physical deployment.

How we built it

We designed Embodied World as a modular world-generation pipeline.

The system begins with a hardware concept generated by CAID Technologies. This may include a textual specification, reference image, system diagram, 3D asset, or structured hardware representation.

The design is then passed through an environment-generation workflow that can produce:

• A panoramic representation of the environment.
• Scene geometry and depth information.
• Gaussian-splat or mesh-based visual assets.
• Semantic information about objects and regions.
• Collision geometry for navigation and simulation.
• A browser-based interactive viewer.

The frontend is designed around modern web-based 3D rendering technologies, allowing users to explore generated environments directly in the browser.

The backend manages world-generation jobs, asset storage, scene metadata, exports, and the connection between the original hardware design and its generated environment.

We also designed the architecture to support multiple world-generation providers. This allows Embodied World to use external world models today while gradually replacing individual stages with open-source or internally developed systems.

Challenges we ran into

One of the largest challenges was determining what a generated “world” actually needs to contain.

A visually convincing image is not enough. An interactive environment must also preserve geometry, scale, camera position, collision boundaries, object identity, and spatial consistency.

Another challenge was balancing explicit 3D reconstruction with generative simulation. Traditional 3D environments are stable and measurable, but they can be expensive to construct. Generative video models can create highly realistic scenes, but they often struggle to preserve the same objects and locations over long periods.

We therefore had to think of Embodied World as a hybrid system rather than a single model.

File formats were another challenge. Gaussian splats, point clouds, meshes, panoramic images, collision models, and streamed level-of-detail assets all serve different purposes. Designing a pipeline that could move between formats such as PLY, SPZ, RAD, GLB, and PNG required careful separation between editable source assets, runtime assets, and physics assets.

We also had to keep the scope realistic. Building a general-purpose world model from scratch is a major research effort, so we focused on designing a product and architecture that can deliver value before every model component is fully proprietary.

Accomplishments that we're proud of

We developed a clear architecture for connecting generative hardware design with world generation.

Instead of treating a generated machine as a static image, Embodied World treats it as an object that belongs inside a larger operational environment.

We are proud that the system was designed around persistent spatial structure rather than only visual generation. This creates a path toward real use cases involving robotics, manufacturing, digital twins, simulation, and embodied AI.

We also designed Embodied World to remain provider-independent. The platform can integrate with existing world-generation systems while maintaining its own scene schema, asset pipeline, and user experience.

Most importantly, we created a broader vision: moving from generating individual machines to generating the environments in which those machines can exist, operate, and be evaluated.

What we learned

We learned that generative design and world generation are deeply connected.

A machine cannot be fully understood without context. Its environment determines how large it feels, how it is accessed, how it moves, what it interacts with, and whether the design is operationally realistic.

We also learned that no single representation is sufficient for every part of an interactive world.

Gaussian splats are effective for visual realism. Meshes are useful for physics and interoperability. Scene graphs provide structure. Semantic data provides meaning. Neural models provide dynamic behavior. A practical world engine must combine these systems rather than choose only one.

Another important lesson was that persistent worlds and generated video are different technical problems. A complete solution needs both stable spatial memory and generative flexibility.

What's next for Embodied World

The next step is to build a working end-to-end prototype that allows a user to generate a CAID hardware design, place it inside an environment, and explore it in the browser.

Planned features include:

• Text-to-world and image-to-world generation.
• Importing generated "Caid" structured hardware designs.
• First-person and third-person exploration.
• Gaussian-splat and mesh rendering.
• Collision and basic physics.
• Object selection and semantic scene understanding.
• World editing and environment expansion.
• Exporting worlds as standard 3D assets.
• Simulating robots and autonomous agents inside generated environments.
• Comparing multiple versions of a generated machine in the same world.

The long-term goal is to make Embodied World a general-purpose environment for the next frontier AI in physical world.

We want users to move from asking, “What should this machine look like?” to asking, “How will this machine exist, operate, and interact with the world?”

Built With

• claude
• codex
• ling-bot
• lovable
• mcp