Download files
Abstract
This thesis contributes to the growing research of multi-agent autonomous systems concerned with the real-time planning. Over the years, autonomous systems consisting of mobile agents have proven to be efficient, robust and versatile tools for exploration (e.g. space robots), military (e.g. for search and rescue operations), and industrial applications (e.g. Google self-driving cars). With autonomous technologies getting matured day by day, deploying multiple autonomous agents to complete complex tasks are getting lot of interests for many different applications. If a single autonomous agent can complete a job, multiple autonomous agents could potentially complete the job faster. However, introducing multiple agents make the overall system more complex since agents now need to be capable of collaborating with each other effectively. Randomly introducing autonomous agents without an effective mechanism for collaboration might negatively impact the productivity.
This thesis is motivated by the goal of making multi-agent autonomous system ubiquitous for real-world applications. We have taken a bottom-up approach in developing algorithmic machineries to address the challenges on our way to satiate that goal.
For a mobile agent operating in a dynamic environment, the success of executing a task hinges on how effectively it can navigate to the target location. Path planning becomes even more demanding if we introduce more autonomous agents in the environment since the agents now have to treat each other as dynamic obstacles. The path planning algorithm not only needs to avoid obstacles but also need to be fast enough to re-plan if the mobile agent encounters unexpected obstacles during navigation. Moreover, the path planning algorithm needs to guarantee that the agent can traverse the path while satisfying its mechanical constraints.
A Simultaneous Replanning Vectorized Particle Swarm Optimization (SRVPSO) algorithm based on stochastic optimization is developed to find out an optimal cost path by avoiding static and dynamic obstacles. The proposed algorithm reduces the computational time of path planning by applying simultaneous replanning strategy. The SRVPSO algorithm is also able to work under some vehicle constraints, such as vehicle size and steering angle. Furthermore, a methodology for traversability assessment of different terrain is developed for risk-free and robust navigation in an unknown environment, while optimizing its total cost.
An effective mission planner is needed for an autonomous system consisting of a swarm of mobile agents to successfully complete a set of tasks. The challenge of the mission planner is to determine the optimized number of agents and associated tasks for each agent for a mission. A Compromise View (CV) model and a Nearest-Neighbour Search (NNS)-based model is developed to address the task decomposition and task distribution problems during mission planning for a multi-agent autonomous system. These models are shown to be highly effective due to their reactive management structure to fulfil their mission successfully. The NNS model can effectively tackle the break down of agent(s). It also has the task switching capacity.
A multi-objective optimization framework for mission planner that determines the number of agents is required for a mission. The mission planner utilizes the developed task decomposition approach to minimize the time to complete a mission as well as the number of agents. The output of the multi-objective framework is a Pareto-optima which is then taken as an input to the decision-making framework where the optimized number of agents is determined based on some user defined constraints and priorities. In measuring the time to complete a mission, the mission-planner utilizes the previously developed path-planner to simulate the trajectories of the agents in navigating the environment to provide the most accurate estimation.
However, an ongoing mission can be effected by unexpected events (e.g. some weather events, unexpected maintenance requirements for the agents etc.). The planning of future missions’ hinges on the ongoing missions since that provides an estimate of the availability of the resources. A realistic forecast model is needed which can utilize the information about the past missions to provide a statistical estimate about the completion of the current missions.
An artificial neural network-based forecasting model is developed to predict the completion time of a mission based on the information about the past missions. The forecasting model is intended to act as a guide for the potential mission planner. Using this numerical model, a future planner can predict the required resources without going through the optimization process.
All these aforementioned algorithmic tools were demonstrated with extensive simulation results and real-time experiments.