Skip to main content

Factor Graph / World Model Lab

Learn one visual language for four systems: GraphSLAM, energy-based models, latent world models, and action-conditioned video simulators.

Factor graph / world model lab

When loop closures really anchor the map

Raise the amber loop factor and blue-green measurement coverage while watching whether the latent pose chain snaps into a globally consistent solve.

GraphSLAM
x₀x₁x₂ℓ₁priorodomodomloopobsobsu₁u₂cz₁z₂

Choose the system

Try a preset

Try this

Start with a drifting chain, then raise loop closure while coverage stays healthy. The whole map should share the correction instead of leaving drift trapped in the last segment.

Watch for

Observed circles and rectangular factors stay fixed. Only the hollow pose and landmark nodes are the geometry the solver is trying to recover.

Diagram language

filled circle = observed variable or chosen action
hollow circle = latent variable the system must infer or roll forward
rectangle = factor / energy term
rounded-side box = forward-computable function

How to read it

Observed circles feed local constraint factors. Only the hollow pose and landmark nodes move during the solve.

Current math lens

Diagnosis

Well-anchored sparse solve

The graph has both local measurement support and a strong long-range revisit. That is the sweet spot: the optimizer can distribute correction over the whole map instead of just polishing local edges.

ML translation

High measurement coverage plus a strong loop closure turns the information matrix from nearly chain-structured into globally informative.

Where this matters

This is what mature pose-graph backends aim for when they aggressively validate loop closures before adding them.