High-level programming support and collision avoidance control for heterogeneous mobile robots

Abstract

Mobile robots with diverse capabilities are being widely adopted in various domains, including search and rescue scenarios, autonomous warehousing, and manufacturing. As these applications evolve and become increasingly complex, the demand for e­fficient programming support and high-performance task execution becomes even more crucial. Traditional programming support often struggle to adapt to the diverse ca­pabilities of robots, including differences in physical design, sensing abilities, and mobilities. Meanwhile, ensuring high-performance task execution faces challenges in terms of perception uncertainties and large-scale deployment. In such contexts, obstacles and potential collisions among robots pose a significant hurdles to safe and efficient cooperation. This thesis studies how to realize efficient programming support and ensure high-performance task execution for heterogeneous multi-robot systems. In particular, we address these challenges through the development of a high-performance middleware for heterogeneous multi-robot systems. To reduce the burden on developers from programming everything from scratch, we present a novel middleware as the programming support, which offers a graph-based programming abstraction and a runtime kernel specifically designed for heterogeneous multi-robot systems. This abstraction simplifies the expression of cooperative missions while the runtime kernel intelligently manages the heterogeneous robots. Meanwhile, we introduce a team-level programming abstraction and a manipulator-level plugin mechanism to facilitate programming for robotic manipulators. These innovative features simplify the programming process, making it more accessible and efficient for users to develop robotic applications. In pursuit of efficient and effective task execution, we introduce a controller-based collision avoidance algorithm that models the perception uncertainties and integrates them into the design of action controller. Besides, to mitigate the imminent collision in environments with a large number of robots operating within a limited space. We present a time-to-collision force-based reward shaping approach to learn a robust collision avoidance policy. This approach facilitates the learning of a robust collision avoidance policy, enabling robots to predict and avoid collisions, thus emphasizing safe navigation and motion planning to enhance task execution efficiency.