| Summary: | In recent years, Multi-Robot Systems (MRS) have gained significant interest in research and in industry (Khandelwal and Stone, 2017; E. Schneider et al., 2016; Amato et al., 2015; Alonso-Mora et al., 2015b; Enright and Wurman, 2011). Manufacturers are moving away from large one-size-fits-all productions to more customisable on demand production, which result in smaller and smaller batch sizes. Additionally, in order to be able to increase productivity even further, more and more tasks in the production process have to be automated. To accommodate these changes, industry is facing major shifts in how the products are produced and in particular the role robotic platforms are playing. Previously, robots have mainly been used in a static manner, i.e. performing a singular repetitive task over and over again with high precision and speed. When multiple robots are employed in such a setup, each robot performs a dedicated task, with no interaction with the other robots. While this approach was suitable for large-scale productions, it cannot maintain the same productivity for highly customisable products. Additionally, many tasks in the production process require that the robots are mobile, since they are spatially distributed. One example is for instance retrieving items from different locations in a warehouse. Furthermore, another requirement is that every robot should be able to handle many different tasks and more importantly, many robots should work together in a team towards a common goal. These new requirements introduce various new challenges. As an example, since the robots are mobile, they should be able to perform the tasks alongside the human workers. Likewise, since multiple robots have to work together, a new challenge is to coordinate such MRS. The work presented in this thesis focuses on the core issues when deploying MRS in the physical world. We focus on the task of warehouse commissioning as a running example. The environment for this task is highly dynamic, adaptive and complex, since new orders can appear at any time and priorities might change. A major issue is to coordinate the robots, while taking current and possible future tasks into account. One solution is a centralised planning entity, which knows about all tasks and robots in the team and assigns the tasks accordingly. While in the case of a handful robots, a good assignment can usually be calculated in a straight forward manner, a problem with a centralised system arises when more and more robots are added to the system. The number of possible assignments rises exponentially with every additional robot. Thus, planning times increase and it might become infeasible to provide an optimal plan in time or to respond quickly to changes. On the other hand, in a decentralised solution, each robot decides on its own. Thus, it accumulates all necessary information, and calculates a plan based on this information. While the robots might not have all information available, this is in many cases not necessary. The planning robot is mainly interested in its own actions. While the robot should take the other robots into account, this effect can be approximated, and not every single action of the other robots is needed. This results in a much less complex planning problem, which allows the robot to re-plan online, as soon as the environment changes. In this thesis, we focus on such decentralised solutions for MRS that can run online on the robots. We investigate navigation, decision making and planning algorithms that are suitable for problems in which the tasks are highly dynamic and spatially distributed, such as the warehouse commissioning example. We explore how a team of robots can navigate safely in a shared environment with humans. We apply Monte Carlo sampling techniques and trajectory rollouts as used in the commonly used Dynamic Window Approach (DWA) (Fox et al., 1997), while taking the localisation uncertainty into account. We show that our resulting navigation method is robust and able to run decentralised on the robots. To facilitate formal evaluation of planning and decision making algorithms, a formal framework called Spatial Task Allocation Problems (SPATAPs) is introduced, that enables us to capture and analyse these problems in the well known Markov Decision Process (MDP) (Puterman, 1994) and Multi-Agent Markov Decision Process (MMDP) (Boutilier, 1996) frameworks. The commonly used MDP solution methods, i.e. value iteration and dynamic programming, fail to provide a solution, due to the large problem space. We investigate whether we can exploit the structure of these problems and introduce approximations to enable planning using the common solution methods. We further refine the framework to formally capture the warehouse commissioning task. A solution method based on Monte Carlo Tree Search (MCTS) (Kocsis and Szepesvári, 2006) is introduced, using computationally cheap greedy roll-out strategies. We show that the resulting approach can yield significantly higher performance than previous approaches, while still being able to plan within the magnitude of seconds, which allows for online re-planning on the robots. Finally, the decision making algorithm and the navigation approach are combined in a proof-of-concept application, in which three youBots are used in a physical warehouse commissioning setup.
|
|---|
