Towards Safe and Resilient Autonomy in Multi-Robot Systems

Abstract

Autonomous systems are envisioned to increasingly co-exist with humans in our daily lives, from household service to large-scale warehouse logistics, agriculture environment sampling, and smart city. Among them, networked cooperative systems such as autonomous multi-robot systems have been widely studied given their capability of accomplishing complex tasks through cooperative behaviors. Reliable interactions among robots as networked safety-critical systems often require provably correct guarantees about safety (e.g. collision avoidance) and resilience (e.g. capability of maintaining communication and operating in an unknown environment). As we strive to design and control such a large-scale system, robots are often assumed to have perfect information (e.g. ground-truth state, system dynamics, and environment model information), unconstrained inter-robot communication, and fault-free operation. However, the precomputed guarantees based on these assumptions could be easily broken when deploying robots in the real world that is uncertain, rapidly changing, and inherently stochastic.

In this thesis, we seek to develop and validate mathematically grounded algorithms to assure safe and resilient interactions among robots that adapt to uncertain and possibly hostile dynamic environments. To achieve the design objective, we discuss three research topics, including (1) safe control and learning under uncertainty, (2) resilient multi-robot interaction through networking, and (3) data-driven multi-robot coordination adapting to the unknown environment.

For (1), we first propose a reactive safe control framework for multi-robot systems under emph{known} robotic system dynamics with localization and motion noise. The framework generates multi-robot motions through centralized or decentralized computation to formally satisfy the collision-avoidance with lower bounded probability guarantee, while respecting the original robot behaviors for task efficiency. When the robotic system dynamics is emph{unknown} and emph{partially} modelled, we then develop a emph{sample efficient} safe learning framework for control that allows the robot to locally learn the unknown dynamics online while achieving sample efficiency in optimizing task performance with bounded regret and safety guarantee.

For (2), we design emph{provably correct} connectivity control frameworks utilizing the graph theoretic and control theoretic approaches for a team of robots to satisfy various global and local interaction network requirements while progressing towards mission goals. This allows the robot team to maintain, recover, or enhance user-defined network connectivity for smooth information exchange under possible adversaries and minimally deviate from their original behaviors. The proposed frameworks prove to be emph{minimally restrictive} towards maintaining the robots' original task-prescribed controllers subject to the connectivity requirements, thus optimally balancing between the required network redundancy and task performance.

With the assurance of safety and resilience in terms of retaining integrity of multi-robot systems through reliable networking, for (3) we address the problem of data-driven multi-robot coordination in the application of sensor coverage to achieve resilient cooperative behaviors in unknown environments. Specifically, a learning-enabled multi-robot control framework is proposed for robots to explore in the unknown environment and simultaneously optimize the task performance regarding sensor coverage using environment model learned online. This allows the robots to share locally collected data or model-related parameters through connected multi-robot network to learn a global environment model and develop task-related coverage controllers with this model to optimize the coverage performance.

Our approaches of safe control and resilient networking for multi-robot systems share a unified optimization-based control framework in real time, thus enabling the synthesis of certified control modules for different task-prescribed multi-robot behaviors with safety and connectivity guarantees. With our approach of sample efficient safe learning for control, we further extend the model-based safety analysis to partially modelled dynamical systems with learning-based behaviors, which enables strong synergies between learning and control with safety guarantee. The integration of data-driven methods into cooperative multi-robot control facilitates the design of adaptive multi-robot coordination behaviors for improved performance in unknown environments. The effectiveness of the proposed methods is demonstrated and evaluated through a set of simulations and realistic simulated robotic platforms.