Advising Agent for Supporting Human-Multi-Drone Team Collaboration

TL;DR

This work tackles the challenge of limited autonomous capabilities in human-in-the-loop multi-drone teams for SAR missions by introducing an advising agent that provides context-aware, non-binding action recommendations. It combines offline reward estimation from a small set of human demonstrations with synthetic trajectory generation to train a reward model that predicts long-term effects of actions, and an online policy (action generator and ranking model) to present top actions. Through extensive SAR simulations and human-in-the-loop experiments, the approach yields higher target detection and area coverage, with agents being well accepted by operators and not increasing cognitive load. The methodology demonstrates that high-quality, long-horizon guidance can be achieved with limited demonstrations, enabling more scalable and effective human-multi-drone collaboration in dynamic, uncertain environments. The core ideas are formalized via an MD P $<S,A,P,R,\gamma>$, a two-phase offline-online framework, and a rigorous QA process to ensure synthetic data fidelity.

Abstract

Multi-drone systems have become transformative technologies across various industries, offering innovative applications. However, despite significant advancements, their autonomous capabilities remain inherently limited. As a result, human operators are often essential for supervising and controlling these systems, creating what is referred to as a human-multi-drone team. In realistic settings, human operators must make real-time decisions while addressing a variety of signals, such as drone statuses and sensor readings, and adapting to dynamic conditions and uncertainty. This complexity may lead to suboptimal operations, potentially compromising the overall effectiveness of the team. In critical contexts like Search And Rescue (SAR) missions, such inefficiencies can have costly consequences. This work introduces an advising agent designed to enhance collaboration in human-multi-drone teams, with a specific focus on SAR scenarios. The advising agent is designed to assist the human operator by suggesting contextual actions worth taking. To that end, the agent employs a novel computation technique that relies on a small set of human demonstrations to generate varying realistic human-like trajectories. These trajectories are then generalized using machine learning for fast and accurate predictions of the long-term effects of different advice. Through human evaluations, we demonstrate that our approach delivers high-quality assistance, resulting in significantly improved performance compared to baseline conditions.

Figures (8)

• Figure 1: The advising agent design. $s$ denotes the state, $a_i\in A$ denotes an action, $s'$ denotes the expected resulting state and $v_i$ denotes the predicted reward from the transition.
• Figure 2: The SAR User Interface.
• Figure 3: Manual control panel, during handling detection alert.
• Figure 4: The left panel contain four tabs: Drones, Areas, Status and Parameters.
• Figure 5: Orange bars are the average number of true targets that were approved. Blue bars are the number of simulations that were performed in this category.