KnowinWorld · Counterfactual Architecture

01

Spatially-Grounded Latent State

Inputs

Multi-view Observation

Robot State

Task Context

3D World Structure

Latent Statez_t

3D Geometry Object State Robot State

Physics-aware Constraints

△Collision

⌁Contact

↗Reachability

◇Object Permanence

02

Z_{t + 1 : t + H} = F_{θ} (z_{t}, A, C_{phys}, b)

03

Same current state z_t · multiple candidate actions · multiple futures

A1Embodied Geometry Violation

❌

A2Contact Dynamics Failure

❌

A3Constraint-Satisfying Future

✓

04

Score predicted futures against the goal and return failure signals.

Task Progress

Safety Risk

Physics Consistency

Reachability

Prediction Uncertainty

Goal-Aligned Ranking

Rank candidate futures by overall alignment with the task goal.