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

TL;DR

This work tackles active search and tracking of multiple unknown dynamic objects with a team of heterogeneous agents and untrusted external information sources. It introduces a time-varying, weighted belief representation, an LSTM-based trajectory predictor integrated into long-horizon planning, and a fully integrated information-driven optimization pipeline coordinated via a central HQ (with robust asynchronous fallbacks). Through extensive Monte Carlo simulations, 3D robotics simulations, and CFD-derived real-world trajectories, the approach demonstrates 1.3–3.2x faster mission completion and improved tracking performance, even when targets outnumber agents by up to 5-to-1. The results highlight the practical potential of combining heterogeneous information fusion, learned intent prediction, and multi-criteria optimization for scalable, real-world multi-agent search and tracking in dynamic environments.

Abstract

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

Figures (26)

• Figure 1: Example scenario: cooperative agents $a_1, a_2$ that search and track unknown target objects $o_1, o_2, o_3, o_4$ in a known environment. The headquarters collects information sources, including agents and independent external report information sources, and coordinates tasks to maximize information gain. $a_1$ is searching the environment following a trajectory with the best utility (blue) on the frontier, integrating the external report. $a_2$ detected $o_3$ and is tracking to reduce $o_3$ location's uncertainty, exploiting a long-term trajectory prediction. Once the uncertainty is below a specified threshold, the target is cleared and does not need to be tracked anymore.
• Figure 2: Simplified example over a small grid that shows: (1) Environment, with two agents ($a_1$, $a_2$ in blue) with their sensor ranges (dashed circles), one object ($o$ in red), and one event of third-party reporting ($e$ in yellow circle); (2) Initial belief without any prior, thus all cells unknown; (3) at time $t_0$ the shared belief after the initial sensor measurements by both agents with the third-party reporting and the frontiers ($f_0, f_1, f_2)$; (4) at $t_1$ the agents move to the optimal frontier locations and get corresponding sensor measurements, where $a_1$ also detects an object $o$ with uncertainty (red dotted circle) and estimates its trajectory (red dashed line); and (5) at $t_2$, the agents select their next action, resulting in $a_1$ achieving the tracking of the object, i.e., reducing the uncertainty (green dotted circle) and clearing the target (green), and $a_2$ continuing independently on the search. Note that the cells (in pink) represent the time-varying object occupancy proposed in this study where the lighter color, the more recently explored.
• Figure 3: Our proposed multi-object active search and tracking system. (a) Overall system architecture for search and track process, with multiple agents that can move and sense, forming a belief with time-varying uncertainty and including third-party reporting, as well as intention-aware tracking, to then select the best trajectory based on both exploration and exploitation values during both modes. (b) A single agent's perspective: during the search phase, to quickly detect the target, the agent follows an optimal trajectory (blue line) towards a frontier, guided by the $J_\mathit{explore}$ and $J_\mathit{exploit}$ utilities that include information from third-party reporting. Once the target is within the agent's visual range, $\mathit{vr}_a$, the tracking phase begins. To rapidly reduce the uncertainty of the target’s location, the agent follows a long-horizon optimal trajectory (blue line), assisted by $J_\mathit{explore}$ and $J_\mathit{exploit}$ utilities that include intent information from learning-based trajectory prediction (pink line). After the task on the current target is completed, the agent searches for the next target. Note that the technical contributions in our paper are highlighted in $C1$ to $C3$.
• Figure 4: Example of cell time-varying uncertainty: (left) case (1) without any target detected; (right) case (2) target detected but lost. For both cases, large $r$ value (dashed) represents a faster convergence than smaller $r$ (solid).
• Figure 5: Environments used for simulation: (left) open$\sim$50 x 50m, (mid, right) city1---Boston, city2---Shanghai $\sim$100 x 100m.