A distributed control framework for motion coordination of teams of autonomous agricultural vehicles

A distributed control framework intended to coordinate the motions of teams of autonomous agricultural vehicles operating in proximity is presented; master–slave and peer-to-peer operation modes are supported. Each vehicle has a nonlinear model predictive tracking controller, which keeps it as close as possible to the path demanded by the task, and coordinates and avoids collisions with nearby vehicles. To do this, it receives the motion trajectories of all other vehicles in its vicinity that may interfere with its own projected motion via a wireless network, and incorporates these trajectories in the computation of its own optimal control action. Each controller is supervised by a higher-level task controller that determines a limited set of the controller's parameters. Simulation experiments have shown that the minimisation of the tracking error along a finite horizon enabled the controller to track paths containing sharp turns, by applying appropriate steering well in advance of the turn. It has also been shown that the variation of specific predictive controller parameters results in a wide range of behaviours; i.e., vehicles can move in both operating modes, coordinate with nearby vehicles by altering their velocity profiles or the shapes of their paths, and avoid collisions.

Introduction

The increasing need for financially competitive and environmentally sustainable agricultural production, combined with a decreasing labour force across agriculture, are the major thrusts towards incorporating information technologies, automation and robotics in the production process. Much effort has been made towards developing autonomous robots for labour intensive agricultural operations such as weeding (Bakker, van Asselt, Bontsema, Müller, & van Straten, 2010; Slaughter, Giles, & Downey, 2008) and harvesting (e.g., De-An, Jidong, Wei, Ying, & Yu, 2011; Kanae, Tateshi, Akira, & Junichi, 2008; Van Henten, Van't Slot, Hol, Van Willigenburg, 2009). The incorporation of robots in many real-world agricultural production processes necessitates the cooperation and coordination of more than one autonomous machine. For example, robotic vegetable harvesting in a greenhouse without a conveyor system would require the cooperation of autonomous harvesting mobile robots with robotic transport vehicles which will move the harvested produce towards certain collection points. Another example is autonomous grain harvesting with on-the-go unloading. This operation requires carefully controlled, coordinated motions of harvesters and transport wagons, so that the produce can be loaded and carried out of the field fast, safely, and without any collisions.

Cooperative mobile robotics research falls into two major categories: deliberative model-based approaches that rely on planning and execution, and swarm type approaches, where cooperative operation emerges from the local interactions between simpler robot behaviours and their environment (Parker, 1998; Werger, 1999). Deliberative approaches address coordination on two levels: mission planning and execution. On the mission planning level, the issues of task allocation, scheduling, and conflict-free optimal global routing must be addressed. In agriculture, examples of recent relevant work include a centralised motion coordination scheme for autonomous tractors performing peat moss harvesting (Johnson, Naffin, Puhala, Sanchez, & Wellington, 2009) and optimal routing for autonomous harvesters and transport wagons (Bochtis, Sørensen, & Vougioukas, 2010). On this level, coordination is global, that is, the planner must take into account all vehicles, tasks and workspace constraints. On the execution level, coordination mostly involves issues of formation control and resolution of local spatial conflicts. On this level, coordination applies to nearby vehicles only and, therefore, uses local information. This paper addresses the problem of coordinated motion control for teams of agricultural robotic vehicles on the execution level. In particular, it presents a distributed control framework in which coordinated motion is achieved by higher-level task controllers residing in each robot. Those task controllers manipulate a limited and well-defined set of parameters in the team's lower level nonlinear model predictive tracking controllers. The framework supports two major modes of operation: i) master–slave, where one vehicle (the master) specifies the motion characteristics of one or more machines (the slaves); and ii) peer-to-peer, where each machine follows its own path while avoiding collisions with other machines. Master–slave operation is required for cooperation, i.e., the coupled, synergistic function of robots that share the same goal. An example of a master–slave operation is on-the-go unloading during grain harvesting, where an autonomous combine harvester (master) specifies the motion of an unloading truck (slave). Peer-to-peer operation is needed for coordination, i.e. the resolution of spatial conflicts when robots operate independently in the same field. An example of peer-to-peer operation is collision avoidance during turning at the headlands of a field, in which many robotic vehicles are operating concurrently. The framework could apply to autonomous vehicles as well as conventional agricultural vehicles operating in supervised auto-guidance mode.

In the general robotics literature, research on mobile robot coordination on the execution level has mainly focused on formation control, i.e., how a group of robots can be controlled in a coordinated way to establish and maintain a formation of a certain shape. The focus is mostly on large or very large teams of similar robots with applications in surveillance, search, and mobile sensor networks. In general, the approach that has received the most attention is decentralised formation control. In distributed behaviour-based approaches (Balch & Arkin, 1998; Carpin & Parker, 2002; Fredslund & Mataric, 2002; Lawton, Beard, & Young, 2003), each robot has several basic behaviours and the control action is a vector weighted average of all behaviours. Behaviour-based approaches have been combined with potential fields (Reif & Wang, 1999; Savvas, Loizou, Tanner, Vijay, & Kyriakopoulos, 2003). Bio-inspired approaches rely on rules based on local sensing to achieve emergent coordination and are based on swarming and flocking behaviours observed in nature (Jadbabaie, Lin, & Morse, 2002; Mondada et al., 2004; Reynolds, 1987). Formation control using Liapunov theory has also been used for groups of robots with non-holonomic dynamics (Mastellone, Stipanovic, Graunke, Intlekofer, & Spong, 2008).

In the context of agriculture, researchers have focused on the master–slave approach to coordination. Noguchi, Will, Reid, and Zhang (2004) proposed a master–slave architecture for two robots performing farm operations. The master commands the slave to either follow a path parallel to it at a fixed distance and angle from it, or go to a certain point along any path, as long as it does not collide with the master. A sliding-mode path-tracking controller was used for steering and velocity control, and collision avoidance was achieved by a risk function based on the master–slave distance which was included in the velocity and steering laws of the slave. Hao, Laxton, Benson, and Agrawa (2004) focused on motion planning for a harvester and grain cart system and used nonlinear control for trajectory tracking. Zhu et al. (2009) developed a master–slave control system for two-tractor platooning along straight and curved paths. Based on the vehicle kinematic model, a reference course for the following tractor was dynamically created by the trajectory of the leading tractor together with a given lateral offset. A linear and quadratic optimal regulator (LQR) path-tracking controller was designed to guide the following tractor along the reference course. However, only the lateral offset between the two tractors could be controlled, whereas the longitudinal distance between the two vehicles was not controllable. Zhu et al. (2009) used a fuzzy logic steering controller to achieve two-tractor platooning on sloping terrain. The reference course for the following tractor was dynamically created from the position points of the leading tractor. Zhang, Geimer, Grandl, and Kammerbauer (2009) implemented a master–slave platooning approach, where the position points of the leading human-operated tractor provide the target position points for the guidance of the following autonomous agricultural vehicle. The path segment between the actual position of the following vehicle and its desired position was represented by a spline function. The follower performed speed and steering control via two independent cascade controllers with feed-forward control. Obstacle avoidance was achieved manually, by issuing a “Lane-Change” manoeuvre command. Zhang, Geimer, Noack, and Grandl (2010) developed this approach further by using a state space dynamic model and a proportional-derivative controller with state feedback and disturbance feed-forward for the tractor. Leader-follower patterns have also been investigated in the general robotics literature (Das et al., 2002; Desai, Ostrowski, & Kumar, 2001; Egerstedt, Hu, & Stotsky, 2001; Vidal, Shakernia, & Sastry, 2004). Sliding-mode control (Sanchez & Fierro, 2003), model predictive control (Wesselowski & Fierro, 2003), and receding horizon control (Dunbar & Murray, 2006) have also been proposed for tracking leader-follower schemes.

In mechanised agriculture, coordinated motion involves relatively small teams of heterogeneous vehicles, (e.g., harvester and truck), rather than large teams of similar vehicles. Furthermore, some operations (e.g., harvesting a field parcel) are performed many times, in exactly the same manner, under similar conditions. Consequently, small inefficiencies during an individual operation can lead to large accumulated inefficiency due to repetition. Hence, optimality of motion coordination is desired in terms of path accuracy, dead distance, time, fuel or other efficiency criteria. This paper explores the utilisation of distributed model predictive control (MPC) as a framework for optimising the coordinated motions of heterogeneous vehicles. More specifically, given desired position and orientation trajectories for all the vehicles, each MPC tracking controller computes optimal control inputs (steering and velocity) that on the one hand, minimise the vehicle's total predicted position and orientation tracking error with respect to its desired trajectory, while on the other they avoid collisions with other vehicles.

The rest of the paper is structured as follows: section 2 presents a hierarchical tracking controller for individual vehicles based on nonlinear model predictive control. In section 3, the coordination framework and two operating modes are illustrated, based on communication among the tracking controllers. In section 4, experimental simulation results are presented. Finally, the conclusions from the experiments are discussed in section 5.