Active Next-Best-View Optimization for Risk-Averse Path Planning
2510.06481v1
cs.RO, cs.CV
2025-10-10
Авторы:
Amirhossein Mollaei Khass, Guangyi Liu, Vivek Pandey, Wen Jiang, Boshu Lei, Kostas Daniilidis, Nader Motee
Abstract
Safe navigation in uncertain environments requires planning methods that
integrate risk aversion with active perception. In this work, we present a
unified framework that refines a coarse reference path by constructing
tail-sensitive risk maps from Average Value-at-Risk statistics on an
online-updated 3D Gaussian-splat Radiance Field. These maps enable the
generation of locally safe and feasible trajectories. In parallel, we formulate
Next-Best-View (NBV) selection as an optimization problem on the SE(3) pose
manifold, where Riemannian gradient descent maximizes an expected information
gain objective to reduce uncertainty most critical for imminent motion. Our
approach advances the state-of-the-art by coupling risk-averse path refinement
with NBV planning, while introducing scalable gradient decompositions that
support efficient online updates in complex environments. We demonstrate the
effectiveness of the proposed framework through extensive computational
studies.
Ссылки и действия
Дополнительные ресурсы: