Navigating the Proximity-Safety Balance: Constraint Decomposition for Human Following in Pedestrian Crowds

What the math gives to ML

The paper contains a useful online uncertainty-calibration mechanism rather than merely a robotics-specific controller. Its adaptive conformal inference update maintains several horizon-specific error radii and increases a radius whenever the realized prediction error exceeds it, while decreasing it otherwise; this gives an operational target for long-run miscoverage under changing prediction quality. A strong transfer is to use these radii to inflate collision margins or gate actions in a learned trajectory predictor, diffusion policy, or RL policy, avoiding a fixed uncertainty margin that is either unsafe or unnecessarily conservative. The main experimental question is whether online-calibrated margins improve out-of-distribution safety at comparable goal-tracking performance and without destabilizing policy training.