J-LAW demo

Joint localization and action-conditioned world modeling. Explore trajectory loop closure and visual manipulation in one page.

Metric trajectory

ground truthopen-loopJ-LAW● loop closure

What J-LAW is doing

Before the revisit: odometry and latent rollouts drift independently.

At the revisit: the factor graph adds metric and latent loop constraints.

After the revisit: alternating updates couple pose x and actionable latent state z.

Pose RMSE · open-loop—
Pose RMSE · J-LAW—
Latent RMSE · open-loop—
Latent RMSE · J-LAW—

Drift and latent consistency

pose error · open-looppose error · J-LAWlatent error · open-looplatent error · J-LAWvertical marker = revisit / loop closure
← Back to trajectory demo

J-LAW · PushT

A lightweight, self-contained PushT task. The circular pusher moves the T-shaped object into the highlighted goal. J-LAW combines visual state, action prediction, and loop-consistency correction to keep the predicted object aligned.

Interactive tabletop

ground-truth objectopen-loop estimateJ-LAW estimategoal region / final placement

What is being demonstrated?

Open-loop: small action and pose errors accumulate, so the predicted T gradually slips away from the observed object.

J-LAW: the actionable latent predicts the effect of each push while periodic visual consistency updates correct the object pose.

Result: both estimates follow the same manipulation, but J-LAW stays aligned closely enough to place the object in the target.

Object pose error · open-loop—
Object pose error · J-LAW—
Goal distance · open-loop—
Goal distance · J-LAW—

Pose and prediction consistency

open-loop pose errorJ-LAW pose errorlatent/action prediction error