← Back ICRA 2026

Efficient Active Search Via Amortized Path-Integral Policies

Tejus Gupta, Arsh Verma, Raymond Song, David Guttendorf, Conor Igoe, Luis E. Navarro-Serment, Jeff Schneider

PDF

AI summary

Key figure (auto-extracted from paper)

Graph Neural Networks trained via behavior cloning can effectively mimic computationally expensive path-integral search policies, enabling real-time, efficient active search that generalizes across diverse maps.

Active search Graph Neural Networks Behavior cloning Path-integral policies Real-time robotics Amortized inference

Problem

Existing active search methods ignore observations made along a robot's path, leading to suboptimal decisions, while path-integral approaches that account for them are computationally prohibitive for real-time deployment.

Approach

The authors use Graph Neural Networks trained via behavior cloning to amortize the computational cost of path-integral policies, allowing robots to efficiently evaluate information gain along entire trajectories in real time.

Key results

GNN policies match expert path-integral performance with less than 1.5% F-score drop
Achieves 11× to 254× speedup in per-decision computation time
Demonstrates robust cross-map generalization and resilience to object density shifts
Successfully validated in large-scale (75,000 m²) field experiments with an autonomous ground vehicle

Why it matters

Provides a practical, real-time solution for autonomous search and rescue operations in unstructured environments without sacrificing decision quality.

Abstract

This work presents amortized path-integral poli- cies that enable efficient and real-time active search for robotic systems. We model search as an active sensing problem where agents select actions to maximize information about target locations. Unlike previous approaches that only consider in- formation gain at final waypoints, our method accounts for observations along entire paths. To address the computational expense of path-integral policies, we amortize costs through Graph Neural Network (GNN) policies trained via behavior cloning. GNNs provide equivariance to spatial transformations and generalize across diverse maps. We validate our approach through field experiments in a 75,000 m2 forested environment using an autonomous ground vehicle, along with simulated testing. Our experiments demonstrate successful policy amor- tization, cross-map transfer, and improved search efficiency.

Index terms

Search and Rescue Robots Field Robots Planning under Uncertainty