Artificial Intelligence-Enhanced Adaptive Controllers for Resilient Unmanned Aerial and Unmanned Ground Vehicles Cooperative Systems: A Mini Review
TL;DR
This updated mini-review aims to bring together state-of-the-art and peer-reviewed research on using artificial intelligence for enhancing adaptive controllers in air-ground cooperation problems and suggests that robust implementation would require explainable AI, rigorous sim-to-real transfer techniques, standardization of heterogeneous benchmarks, and efficient lightweight algorithms suitable for limited resources.
Abstract
Teams of cooperative unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) are becoming crucial in scenarios where their complementary mobility, sensing capabilities, endurance, and physical interaction are beneficial. While UAVs offer fast aerial inspection and flexible perspectives, UGVs ensure large payloads, extended operation time, and the ability to intervene from ground level. Nonetheless, reliable UAV–UGV cooperation is challenging due to different dynamics, occasional communication, limited computational power, uncertain environments, and severe constraints on energy. This updated mini-review aims to bring together state-of-the-art and peer-reviewed research on using artificial intelligence (AI) for enhancing adaptive controllers in air-ground cooperation problems. This paper addresses four interconnected topics: learning-enabled control, model predictive and hybrid AI-model predictive control controllers, resilient coordination considering limited communication, and energy-conscious mission planning. As opposed to UAV–UGV surveys, this paper highlights the control-related aspects of air-ground cooperation, which include learning policies, constraint satisfaction, task allocation, degraded communication-based re-planning, and energy-dependent rendezvous. The revision significantly improves the process of selecting relevant literature, eliminates claims on benchmarks without evidence, updates unverified citations or references unrelated to the work, and produces publication-quality graphics. Ultimately, the review suggests that robust implementation would require explainable AI, rigorous sim-to-real transfer techniques, standardization of heterogeneous benchmarks, and efficient lightweight algorithms suitable for limited resources.