Decentralized coordination enhanced by centralized information: multiple AGVs in industrial application Valerio Digani, Lorenzo Sabattini, Cristian Secchi and Cesare Fantuzzi I. I NTRODUCTION This research deals with the path planning and coordi- nation of multiple Automated Guided Vehicles (AGVs) in an automated warehouse. Coordinated motion of groups of autonomous vehicles is a relevant topic in the field of mobile multi-robot systems and it has thus been widely studied. The contribution of this work is the definition of a methodology to solve the coordination problem in a holistic manner. In particular we present a partially decentralized control strategy for the coordination of multi-AGVs systems based on hierarchical architecture that implements path planning on a two layer framework. The high-level layer describes the topological relationships among different areas of the environment. In the low-level layer, each area includes a set of fixed routes, along which the AGVs have to move. Each AGV autonomously computes its path, on both layers. The coordination among the AGVs is obtained exploiting shared resources about the state of the fleet (i.e. centralized informa- tion) and local negotiation (i.e. decentralized coordination). Preliminary results on this topic were introduced in [1] and a more exhaustive treatment can be found in [2]. II. PROBLEM STATEMENT The goal of this research is to coordinate a fleet of AGVs by reducing the computational burden and the complexity currently necessary for a centralized solution. To this aim, we propose to exploit a decentralized approach that allows to obtain a computationally cheap and useful, even if sub- obtimal, solution. More precisely, the problem consists of planning a path on a route map for a fleet of AGVs in an 2D static environment, in such a way that conflicts and deadlocks are avoided. Each AGV starts its path from a known initial position and it has to reach its own pick/drop position. Each AGV can communicate with its neighbouring vehicles and it has a prior knowledge of the environment. III. TWO LAYER CONTROL ARCHITECTURE The coordination problem is addressed through a hier- archical approach. In particular, two layers are exploited: the top-layer, or Topological Layer, and the bottom-layer or Route Map Layer. Authors are with the Department of Science and Methods for Engineering (DISMI), University of Modena and Reggio Emilia, Italy {valerio.digani,lorenzo.sabattini, cristian.secchi, cesare.fantuzzi}@unimore.it This paper is written within PAN-Robots project. The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement n. 314193. A. Topological layer The first layer is a topological representation of the real map by means of an undirected graph where each node represents an abstract entity called sector. A sector is a region of the real map which can be distinguished from the other ones based on several features (e.g. topological aspects, logistic aspects, etc.) and it has specific properties (e.g. the number of AGVs currently contained inside the sector, maximum number of operations of loading/unloading, etc.). We assume that each pair of neighboring sectors is con- nected by at least one path on the route map. The properties (number of AGVs, etc.) owned by each sector are stored in a centralized manner in order to be visible to all the AGVs. This shared information is used to plan the sub-optimal route for an AGV to reach its destination minimizing the length of the path. Each AGV autonomously and asynchronously plans its path on the sectors graph and the result is a sequence of sectors to cross. Since the future positions of the other AGVs is unknown in advance, a D* algorithm combined with an MPC (Model Predictive Control) mechanism, is exploited to perform path planning. Each sector is weighted using a time varying congestion related metric (e.g. the number of AGVs in a sector). The planned path is then dynamically updated while the AGV is tracking it. At each step, the path is re- planned for optimizing it on a given receding horizon. This approach provides an optimal local solution but a sub-optimal global one, because only the part of the path inside the horizon is interested by the optimization. The procedure is summarized in Algorithm 1: in this algorithm, the vector path contains the list of sectors in the planned path, the term H is the prediction horizon, and the index i identifies the current step. Algorithm 1: Path planning on the topological layer 1 while path[i] = goal do 2 path[i : i + H]= D*; 3 go to path[i + 1]; 4 end B. Route Map layer Inside each sector the coordination among AGVs is needed. The second layer manages the real path following on the route map and the avoidance of deadlocks and conflicts