Predicting 1.6 s ahead for every person in the ward.
runs logged0
trajectory rows0
CLONED POLICY
-
Behaviour cloning on those (observation, action) pairs.
train ·
held-out loss
drivinghuman
SAME ROBOT, TWO PLANNERS
NAIVE
PREDICTIVE
staff hit
0
0
metres
0
0
Both cover the same ground.
The green one hits fewer people.
WSdrive
ADsteer
draglook around
What was actually trained
Everything below is measured in this
browser, on this run. Nothing is precomputed.
Where does prediction actually help?
Same controller and map in every cell, only the
planning input changes. Stratified by crowd density and corridor geometry, after Liu et al.,
"Beyond ADE and FDE" (arXiv 2510.10086) — an evaluation framework, not a model. The predictor half of
its argument, near versus distant agents, is on the model card.