ACTIVE PROGRAM - FLAGSHIP

Teach models to perceive, reason about, and act in continuous space - not just write about it.

THESIS

Frontier models are fluent in language and shaky in space. They can describe a room and still fail at placing a cup, planning a path, or keeping a coordinate frame stable across turns. Spatial Intelligence gives agents geometric grounding: occupancy, affordance, egocentric and allocentric frames, and a world model that updates as the scene changes.

THE PROBLEM SPACE

01Language-only agents invent geometry: distances, occlusions, and left/right flip under mild rephrasing.

02Digital twins, warehouses, and facilities ops need agents that share coordinates with CAD, BIM, and robotics - not guesses.

03Multimodal models that 'see' often treat space as a one-shot description instead of a state to keep and query.

04Without spatial memory, multi-step physical or simulated tasks collapse into re-describing the scene every turn.

APPROACH

World-model loops: keep an explicit spatial state (maps, object poses, free space) that language conditions on - and that tools can read/write.

Align text, imagery, depth/point clouds, and structured scene graphs.

Translate between egocentric and allocentric frames so an agent can reason from a camera view and still plan on a shared map.

Eval suites that punish fluent nonsense: metric error on pose, path validity, occlusion reasoning, and long-horizon consistency.

Wire spatial actions (move, place, annotate) into Agent Cloud governance the same way as any other tool call.

ARCHITECTURE SKETCH

How the pieces fit

01

SPATIAL STATE STORE

Versioned scene state - meshes, bounding volumes, semantic labels, confidence - queryable by agents and humans.

02

PERCEPTION ADAPTERS

Ingest for RGB-D, LiDAR-class clouds, CAD/BIM imports, and synthetic renders from Blender-class pipelines.

03

SPATIAL REASONER

Planner and critic that propose actions in space, simulate outcomes, and abstain when geometry is underdetermined.

04

LANGUAGE ↔ GEOMETRY BRIDGE

Grounds referring expressions ('the pallet behind the yellow rack') to concrete entities in the state store.

OPEN QUESTIONS

  • How much geometry must be explicit vs. latent for reliable long-horizon agency?
  • What affordance abstraction transfers from warehouse to office to virtual set?
  • How do we audit spatial decisions the way we audit text tool calls - with replayable evidence?

WHY IT MATTERS

Agents that can co-pilot facilities, logistics yards, and digital twins without constant human re-description.

Safer automation: policies can constrain spatial blast radius the same way they constrain payment amounts.

A foundation for Blender & 3D Agency - language that can operate a scene graph, not just chat about one.

RELATED PROGRAMS

Book a demo