Child-safety scoring now follows climbing chains: 0.986 rank correlation in simulated homes, 0.947 in classrooms with 25 labels
Synopsis
Many injuries to young children come from relations between objects, such as a chair beside a sofa under a window. AuroraNet turns a room into an age-scaled climbing graph, sums severity-weighted expected harm along sequences a child can climb, and learns 15 edge parameters from scene-level labels. On simulated rooms it reaches a 0.986 rank correlation with gradient boosting, 0.878 versus 0.798 within strata; all evidence is simulation-based.
Interpretation
The object of child home-risk assessment shifts from individual furniture pieces to sequences of objects a child can climb in succession, with severity-weighted expected harm as the output. Earlier scoring treated risk as a property of a single object or of the whole scene, whereas the question here is about relations such as a chair next to a sofa beneath a window. On simulated rooms the reasoner matches gradient boosting at a 0.986 rank correlation and beats it within strata at 0.878 versus 0.798, though the advantage reverses under functional-form mismatch.
The implementation first converts a metric scene into an age-scaled climbing graph, then accumulates harm along its edges. The learnable version fits only 15 edge parameters from scene-level labels, avoiding manual annotation of individual hazard chains. Evaluation covers 1,000 simulated homes per seed across 10 seeds, plus 600 denser homes for a covariate-shift test and 1,000 simulated kindergarten classrooms per seed across 5 seeds.
Monocular and time-of-flight depth fusion recovers scale and propagates uncertainty, while counterfactuals rank furniture relocations. Without real scale, climbing heights and window-sill heights cannot be compared, so harm probabilities cannot be computed. Transfer to classrooms reaches 0.947 with 25 labels; a graph neural network ranks better with 600 labels but needs eight times more labels to match.
Perspective
The evaluation covers room layouts in simulated homes and simulated kindergarten classrooms, with harm labels from a Monte Carlo simulation of a climbing child that combines window falls and bookshelf or storage-tower tip-overs at given weights. Intended readers are researchers and engineers working on scene understanding, home robotics and child-safety assessment, who would split rooms into training and test by seed and compare rank correlations. The source states no next-step experiment.
All evidence is simulation-based, and the source states there are no human-subject data. The advantage over gradient boosting reverses under functional-form mismatch, and the graph neural network ranks better with ample labels but needs eight times more labels to match. Classroom transfer results likewise come from simulated classrooms.