Uncertainty-guided Next Best Iso-Surface Optimization for Neural Volumetric Capture

Robots deployed for ultrasound scanning, industrial inspection, and subsurface mapping must often recover previously unseen volumetric geometry from sequential cross-sectional measurements. Fixed acquisition trajectories cannot adapt their sensing effort to object-specific ambiguity, while view-centric next-best-view objectives do not directly represent uncertainty inside a sparsely observed volume. We formulate active cross-sectional acquisition as epistemic-uncertainty minimization over an implicit occupancy field and instantiate this formulation in SliceOpt. A deep ensemble represents plausible reconstructions; predictive disagreement defines Ensemble Uncertainty, and boundary-localized classification ambiguity defines Boundary Risk. A plane-level information-gain objective combines these signals with geometry- and topology-aware cues and an exploration penalty, enabling closed-loop selection without an object-specific acquisition trajectory. Across synthetic objects and constrained sensing spaces, SliceOpt improves reconstruction fidelity by up to 52% over the strongest passive baseline. At matched slice budgets, it also yields 42% shorter end-effector travel on average than an adapted view-centric planner, although motion is not optimized explicitly. Deployments on an industrial CT turntable and a UR5e-mounted B-mode ultrasound probe demonstrate online adaptation to unknown physical objects. These results position uncertainty-aware active sensing as a practical complement to zero-shot physical intelligence: general-purpose robots still require closed-loop mechanisms that identify and reduce uncertainty after deployment.