Kairos extends a hierarchical 3D scene graph into a 4D one that forecasts where people will be, and which way they will move, at times the robot has not observed. Each voxel stores a directional mixture and a presence rate, spectral predictors forecast both for any query time, and per-voxel variances give credible intervals that tighten as data accumulate. On a campus dataset, a shopping mall and a station concourse recorded for eleven months, its forecasts compete with dedicated occupancy and flow models while learning from only the fraction a patrolling robot sees, and its state stays consistent under loop-closure corrections.
Relevance
Eleven months of data and a memory that survives loop closures is the kind of evaluation long-term autonomy needs. See dynamic scenes and 3D scene graphs; it shares authors with the 3DSG uncertainty paper.
Links
- arXiv · code
- Subfield: Semantic mapping