Reliable and resource-aware collaborative SLAM for multi-robot search and rescue

Abstract

Collaborative simultaneous localization and mapping (CSLAM) is a core capability for many multi-robot systems operating in GPS-denied environments. Motivated by the application of forest search and rescue, this thesis presents two contributions that advance state-of-the-art CSLAM systems. The first part proposes a reliable loop closure procedure for the forest environment. Complex occlusions and strong perceptual aliasing often make it extremely difficult to detect previously visited areas. By leveraging stable tree features extracted from the environment, our approach significantly improves precision and recall during loop closure detection. The proposed technique is fully integrated into a centralized CSLAM system, and is extensively validated during real-world collaborative exploration missions in the forest. The second part of this thesis proposes a resource-aware framework for distributed loop closure detection in CSLAM. Detecting inter-robot loop closures is a resource-demanding process that involves exchanging observations and verifying potential matches. This poses severe challenges for mobile robots as they are frequently limited by available onboard resources. We propose a principled framework for robots to seamlessly adapt to such resource constraints while maximizing performance. Given budgets on computation and communication, our proposed method maximizes a task-oriented performance metric by selecting and verifying a budget-feasible set of potential loop closures. We show that this problem is NP-hard in general. Then, we provide simple approximation algorithms and leverage results on monotone submodular maximization to establish provable performance guarantees. The proposed framework is extensively evaluated on real and synthetic SLAM benchmarks.