[ Consequences ]

A lack of true 3D understanding handicaps the following real-world use cases:

1. Automating 3D scenes
   [ E.g. architecture visualisation, games design, etc.  ]
   AI struggles to select and position the right assets due to limited object understanding and a lack of spatial reasoning.

2. Embodied intelligence
   [ E.g. robotics  ]
   AI struggles in uncertain environments from a lack of robust understanding of individual objects - their sizes, shapes, positions, properties, and functions.

[ Analogy - What LLMs did for language, we’re doing for 3D ]

Google Translate and ChatGPT use similar algorithms - but Google Translate learned task of translation and gained partial language knowledge. ChatGPT learned the structure and content of language itself, enabling broad generalisation across language tasks.

Similarly, most 3D LLMs, and VLMs gain partial 3D knowledge as a by-product of solving specific tasks. The difference: our model is trained to understand objects, spaces, and their relations - the structure, content, and function of 3D environments - enabling broad generalisation across 3D tasks and industries.

↑ Our early proof of concept [ demo video ← ] revealed the urgent need - and huge opportunity - for spatial intelligence.
_

[ Overview ]

SpaceForm Technologies - a leader in 3D scene generation for architecture visualisation - faced a major bottleneck:

40% of their project time was spent manually selecting and positioning 3D assets - slowing delivery and constraining growth.

Scene composition - i.e. placing the right objects in the right place - was a critical pain point.

They needed intelligent automation to increase project throughput and scale their operations.

[ The Project ]

Over a six-month collaboration, we developed a single-prompt system that intelligently selected and positioned 3D assets - dramatically accelerating 3D scene creation.

[ The Limitation ]

While the proof of concept demonstrated real-world value, it was not productisable:

Current 3D and vision-language models lacked true spatial awareness and reasoning.

The system could understand objects and spaces in 2D, but struggled with spatial awareness and understanding 3D environments.

Advances in vision transformers, foundation models, and self-supervised learning have made it possible to move beyond single-task 3D AI.

At Spatial Intelligence, we're combining these advances with new 3D-centric learning objectives focused on object, spatial, and relational understanding.

The technology is ready - we seek to apply it across tasks, industries, and real-world applications.

By learning about the structure, content, and function of the 3D world - not just generating objects or captioning scenes - we enable broader generalisation across tasks and industries.

Our model's flexibility means one core intelligence can automate scene design, generate simulation environments, and enable robotic reasoning - without retraining from scratch - unlike our competitors. This defensibility compounds as the model scales.

We expect competition - but acting now gives us first-mover advantage, and our novel approach strengthens our moat over time.

We aim to become the default 3D intelligence engine - delivering real-time, generalisable spatial understanding across design, simulation, and robotics.

From [ one-click 3D scene generation ← ] to robotic interaction with novel environments, our model will serve as the foundational perception and intelligence layer.

If executed well and backed properly, there’s no reason any application or industry that needs to understand, construct, or interact with 3D environments - real or digital - is out of reach.