Multi-Object Active Search and Tracking by Multiple Agents in Untrusted, Dynamically Changing Environments

• DOI: 10.48550/arxiv.2502.01041
• Publication Date: 2025-02-02
• Authors: Mingi Jeong; Cristian Molinaro; Tonmoay Deb; Youzhi Zhang; Andrea Pugliese; Eugene Santos; V. S. Subrahmanian; Alberto Quattrini Li
• Keywords: Tracking (education)

This paper addresses the problem of both actively searching and tracking multiple unknown dynamic objects in a known environment with cooperative autonomous agents partial observability. The target ends when uncertainty is below threshold. Current methods typically assume homogeneous without access to external information utilize short-horizon predictive models. Such assumptions limit real-world applications. We propose fully integrated pipeline where main contributions are: (1) time-varying weighted belief representation capable handling knowledge that changes over time, which includes reports varying levels trustworthiness addition agents; (2) integration Long Short Term Memory-based trajectory prediction within optimization framework for long-horizon decision-making, reasons time-configuration space, thus increasing responsiveness; (3) comprehensive system accounts enables information-driven optimization. When communication available, our strategy consolidates exploration results collected asynchronously by sources into headquarters, who can allocate each agent maximize overall team's utility, using all available information. tested approach extensively simulations against baselines, robustness ablation studies. In addition, we performed experiments 3D physics based engine robot simulator test applicability real world, as well trajectories obtained from an oceanography computational fluid dynamics simulator. Results show effectiveness method, achieves mission completion times 1.3 3.2 faster finding targets, even under most challenging scenarios number targets 5 greater than agents.