Multimodal models translate images, point clouds, CAD and language into candidate scene graphs, but their outputs remain hypotheses with calibrated uncertainty. Specialised networks estimate pose and affordance; symbolic solvers enforce exact dimensional constraints; physics engines test stability and collision; planning models search sequences of actions. A critic compares generated plans against geometry at every step and returns a structural explanation rather than a generic failure. Human users can correct perception, lock constraints and inspect alternative paths. No single model is asked to perceive, reason, plan and verify simultaneously, because separating those responsibilities is essential to trustworthy spatial intelligence.
PROJECT 008 · Deep-tech concept
Spatial Reasoning Engine
“Can AI understand space well enough to design with us?”
MOVE TO EXPLORETHE IDEA IN ONE LINEspatial
01 / PREMISE
The beginning
Spatial Reasoning Engine is a geometry-native intelligence layer for systems that need to understand what fits, connects, moves, collides, supports or can be assembled. Language models can describe a room or object convincingly while remaining unreliable about exact position, scale and multistep physical transformation. I imagine an architecture in which perception, geometric representation, constraint solving and planning remain explicitly connected. Natural language can state intent, but measurable space governs what is possible. The system would support machines and people designing, building and acting inside real environments where a small geometric error can propagate into failure.
Spatial work is fragmented across computer vision, CAD, simulation, robotics and textual instruction. Perception systems produce labels without robust physical relations; design software stores precise geometry but has little semantic understanding; generative models create plausible arrangements that violate tolerances, access paths or assembly order. Each industry builds narrow integrations around the same underlying gap. Robotics, construction, warehousing and industrial design need reasoning that preserves uncertainty from sensing, obeys hard constraints and explains failure in terms a human can inspect. Treating spatial problems as additional tokens hides the very structure that makes them difficult.
02 / THE PRODUCT
What it could become
The product would be an API and model family that turns scenes and assemblies into typed geometric graphs. Objects carry shape, pose, affordance, tolerance and uncertainty; edges represent contact, support, containment, clearance, articulation and required sequence. A user can ask whether an assembly fits, how it could be reconfigured or why a plan fails. Neural perception proposes objects and relationships, exact solvers enforce geometry and physics, and planning searches valid transformations. The first product could focus on a bounded 2D or constrained assembly domain with unambiguous evaluation, then expand toward 3D scenes, robotic manipulation and collaborative design.
For whom
- Robotics and automation teams
- Architecture and construction software
- Industrial design and manufacturing
- Warehousing and logistics platforms
Core capabilities
- 3D scene and assembly graphs
- Constraint-aware geometric search
- Physical affordance prediction
- Language-to-spatial-plan translation
The project becomes meaningful only when a new technical possibility is translated into a clear human advantage, an experience people can understand, and a system capable of earning trust over time.
03 / FOUNDATIONS
Intelligence and mathematics
Computational geometry provides collision, containment, visibility and configuration-space reasoning. Rigid-body kinematics and Lie groups describe movement and orientation; topology captures connectivity and passages that survive geometric deformation; constraint-satisfaction and mixed-integer optimisation handle tolerances and discrete assembly choices; graph search discovers feasible action sequences. Probabilistic geometry propagates uncertainty from sensing into downstream plans instead of replacing a noisy measurement with one confident coordinate. The system can calculate margins—how far a plan is from collision, instability or infeasibility—so users see robustness rather than a binary answer. Geometry is not visual decoration here; it is the substrate of intelligence.
04 / VENTURE LOGIC
How it might live
The engine would begin as developer infrastructure for one domain where spatial errors are frequent, expensive and digitally observable. Pricing could follow verified spatial tasks or compute, with domain modules for assembly, logistics, construction or robotics. Early partnerships would contribute real scenes, tolerances and failure cases needed to build credible benchmarks. Long-term defensibility would come from geometry-native training data, solver integration and evaluation suites that measure physical validity rather than linguistic plausibility. Safety-critical uses would require domain certification and deterministic fallbacks. The company should expand only when it can prove that the same core representation transfers without concealing domain-specific risk.
For me, a venture is more than an interesting technology. It needs a narrow first user, a repeated problem, a distribution path, a credible advantage and a reason to improve as more people use it. I would test those conditions before deciding whether this idea should become a company, a product, an open technology or an ongoing research programme.
05 / DESIGN PRINCIPLES
Rules for making it real
- 01
Geometry is truth, not decoration.
- 02
Every plan must survive a constraint solver.
- 03
Represent uncertainty in perception explicitly.
- 04
Explain failure in spatial terms a human can inspect.
06 / ROADMAP
From question to company
- 01Frame
Build a benchmark for spatial constraint reasoning.
- 02Prototype
Release a 2D assembly prototype.
- 03Prove
Expand to 3D scenes and robotic planning.
- 04Build
Develop domain partnerships and proprietary datasets.
07 / HONEST QUESTIONS
What could go wrong
Serious imagination includes the possibility that an idea should change radically—or should not exist. These are the tensions the project would need to resolve:
- Synthetic training geometry failing in messy environments.
- Small perception errors compounding during planning.
- Safety-critical use before adequate verification.