Active Next-Best-View Optimization for Risk-Averse Path Planning
Amirhossein Mollaei Khass, Guangyi Liu, Vivek Pandey, Wen Jiang, Boshu Lei, Kostas Daniilidis, Nader Motee
AI summary
Problem
Autonomous robots tasked with goal-directed navigation in unknown or partially observable environments struggle to simultaneously ensure safety along a constrained trajectory and actively gather informative observations, as existing methods typically decouple exploration, perception, and safety-aware planning.
Approach
The framework constructs conservative risk maps using Average Value-at-Risk statistics on an online-updated 3D Gaussian Splatting map to refine a coarse reference path, while optimizing next-best-view poses on the SE(3) manifold via Riemannian gradient descent to maximize proximity-weighted information gain relevant to imminent motion.
Key results
- A risk-averse replanning framework integrating local A* with real-time AV@R-based safety filtering on 3D Gaussian Splatting maps
- A Riemannian optimization scheme on SE(3) for next-best-view computation that maximizes information gain under geometric and task-specific constraints
- Scalable gradient decompositions enabling stochastic or mini-batch updates for efficient online map refinement
- Extensive computational studies demonstrating effective safe navigation and targeted information acquisition in complex 3D scenes
Why it matters
Enables autonomous robots to safely navigate constrained, goal-directed missions in dynamic or partially known environments by tightly coupling safety verification with active perception.
Abstract
Safe navigation in uncertain environments requires planning methods that integrate risk aversion with active perception. In this work, we present a unified frame- work that refines a coarse reference path by construct- ing 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.