Kairos extends 3D scene graphs to 4D to forecast pedestrian presence and directional flow

zhenjun_zhao · x · 2026-09-24

Kairos extends hierarchical 3D scene graphs into 4D scene graphs, storing per-voxel directional mixtures and presence rates so robots can forecast when and where people will move. Spectral predictors output presence probability plus full directional distributions with calibrated credible intervals. Evaluated on a robot-collected campus dataset and a mall and station concourse recorded for eleven months, it rivals dedicated occupancy/flow models trained on full detection streams despite learning from only the fraction a patrolling robot observes.

Original post →

More from Embodied

Embodied channel →