Collaborative Vision-based Simultaneous Localization And Mapping for Robotic Teams

Abstract

Simultaneous Localization And Mapping (SLAM) constitutes one of the most fundamental problems in robotics, since ego-motion estimation and map-building are key in enabling autonomous navigation. Allowing robots to perform tasks independently, relying on onboard sensing systems only, SLAM forms the basic building block of many applications for mobile robots, such as autonomous vacuum cleaning or package delivery. While most state-of-the-art navigation solutions are designed for a single robot only, robotic collaboration promises increased robustness and efficiency of missions with great potential in applications, such as search-and-rescue or agriculture. Beyond that, algorithms for autonomous navigation form the basis for contemporary Augmented and Virtual Reality applications with mobile devices such as smartphones or head-mounted displays. Multiple robots and other vision-equipped devices, also considered as agents in a multi-agent system, collaborating in a decentralized fashion or centralized by communicating through a server, can substantially boost the efficiency of a mission by dividing up tasks, for example the load of mapping an environment. Furthermore, co-localization of the agents is key for collaborative tasks, such as carrying a load. Sharing information in a robotic team allows avoiding duplicate work, and the access to more than one’s own experiences promises better accuracy of estimates at runtime. With increasing maturity and robustness of single-agent SLAM, multi-robot systems have been gaining growing popularity over the last years. However, such systems are still in their infancy, with only a few works tackling the SLAM problem in a collaborative fashion, often making assumptions imposing substantial limitations to the system, such as perfect network connection or known initial configuration of the agents, or are only tested on pre-recorded datasets or with limited data exchange. The main challenges in collaborative SLAM lie in maintaining a consistent estimate with experiences from multiple contributors, efficient data management to handle the large amount of data collected by all participating agents, and transparency of information throughout the system to exploit the full potential of collaboration. Furthermore, robust data exchange and efficient algorithm design are necessary to enable the applicability of multi-agent SLAM systems in real-world missions. To this end, this doctoral thesis has been investigating collaborative SLAM, focusing on centralized topologies in a vision-based sensor setup. In particular, effective approaches of system architecture for centralized collaborative SLAM have been studied such that data transparency, efficiency and practicality can be maximized. This research has led to the development of two multi-agent SLAM systems, allowing robotic agents to collaboratively perceive their workspace by sharing their experiences of the environment with each other through a central entity. The evaluation on simulated datasets as well as real experiments reveals not only more efficient SLAM estimates in the collaborative case in comparison to the single-agent case as expected, but also more accurate estimates for all agents at runtime, as each agent has access to more than its own data at each instant. Moreover, the computational bottleneck arising from an increasing number of agents contributing with data in the SLAM estimation process is studied, assessing the impact on the scalability of collaborative SLAM. Following this investigation, the last contribution of this thesis aims to mitigate this explosion of data in the system employing well-studied notions in Shannon’s Information Theory to assess the extent of redundancy of information within the SLAM graph before identifying and removing parts of this graph that consume precious computational resources at diminishing returns to the accuracy of the SLAM estimate. The algorithms and systems proposed in this thesis are evaluated in challenging real-world applications as well as state-of-the-art benchmarking datasets, attesting to their practicality and accuracy. While mainly experiments are presented on small aerial robots as a particularly challenging platform for vision-based algorithms, the principles researched in this thesis are applicable to a wide range of platforms equipped with vision sensors. This research opens up new possibilities for multi-agent collaboration and mission control coordinating multiple agents, enabling the team to make the most of each participating agent, for example in a heterogeneous robotic team with agile drones overlooking the scene and ground robots with limited mobility but higher payload and processing power than the airborne robots. Altogether, the contributions of this thesis add to the efficiency, practicality and accuracy of multi-agent SLAM, bringing collaborative SLAM one step closer to real-world deployment