2.3 Experimental Architecture
2.3.2 Decentralized Model
As we transition towards algorithms and experiments that require larger and more complex environments, it is necessary to move towards a truly decentralized testbed while attempt- ing to maintain some qualities of the centralized model. Since the centralized approach to experimental testbed design required a persistant network connection for the purposes of data logging and code deployment, its range is limited to that of a single access point. Switching between multiple IEEE802.11 access points does not occur quickly enough, e.g. on the order of 5 s, for real time control. Thus, in order to operate effectively across the environment in Fig. 2.4 we rely on several enabling technologies and components.
Localization
A second order strength of ROS is a large library of software for basic robotic tasks that has been developed and tested by Willow Garage [106]. The open-source release of the
navigation stack with some optimizations for our platforms provides reliable laser-based localization and motion planning across very large environments as depicted in Fig. 2.4.
Figure 2.4: Map for laser-based localization. Note that under the centralized model, experiments were confined to the range of a single access point. The truly decentralized testbed allows for experiments that range throughout the available environment.
Mesh Network
The open80211s consortium [80] is developing an open-source implementation of the emerging IEEE 802.11s mesh networking standard and distributing it into modern linux kernels. It works effectively to allow transparent multi-hop routing of network traffic be- tween many systems operating on the mesh where mesh routes are dynamically computed based on current conditions. By bridging the mesh network with our local wireless/wired network, we create an environment where laptops and desktops used for development are connected via the mesh network to robots across the building.
Decentralized Operation
Though the mesh network extends in the attempt to provide ubiquitous network connec- tivity, it is inevitable that during some experimentation, the connectivity will fail. In order to be robust to such failures, we seek to deploy ROS in a truly decentralized way. Though the focus of ROS development is on a single robotic system, the PR2, ROS mes- sage passing takes place via UDP or TCP/IP channels and can operate over a network on multiple machines. However, the presence of a central master that manages message connections makes it not well suited to deal with intermittent connectivity.
We have solved this problem by augmenting the default ROS master with a version that performs automatic service discovery of other ROS masters and allows for synchronization of communication channels that are currently being managed. In this way, an experimental system is designed such that each robot locally registers it’s hardware drivers, algorithms, and communication channels with its local ROS master. When mesh connectivity is available, each robot’s ROS master discovers and synchronizes its table of running software and interfaces.
Data Logging
The ROS environment provides an efficient means to log arbitrary messages to a local filesystem. This means that logging of state information, communication messages, and other sensory data on an individual robot is trivial. However, the intermittent connectivity of a truly decentralized experimental testbed complicates the matters of data logging for experimental verification. We address this problem by ensuring that each robot operates on a synchronized clock and aggregating individual logs after each experimental trial. While this does not provide quite the same ability of the centralized testbed where the full state of the system can be queried at any time during an experiment, it is a robust and reliable way to log large amounts of data without overburdening the mesh network.
Chapter 3
Coordinated Control
We present two examples of problem statements targeted for networked teams of robots. The first is a case study involving caging manipulation by a team of agents. It assumes a disc model for local communication with asynchronous broadcast control to facilitate decentralized controllers. The second example is more tightly coupled as it considers the networking and mobility problems associated with providing general situational awareness capabilities to a team of mobile robots. It serves as motivating problem statement for the remainder of this thesis.
3.1
Multi-Robot Manipulation via Caging
The concept of object closure or caging for manipulation is introduced in [82] where mul- tiple robots are used to geometrically confine a payload to a compact set in the plane. Here we extend this method for manipulation with a decentralized algorithm for large teams of robots [27, 29] by employing sequential composition of vector fields as in [42]. Team behaviors are chosen that approach, surround, andtransport a payload where each behavior is defined by a vector field that requires global task parameters and local state information allowing it to operate within the ABC architecture. Transition between be- haviors is accomplished by local sensing and message passing as depicted in Fig. 3.1. In this way we are able to construct a system for payload transport that is robust to agent
addition or deletion, the presence of obstacles, and a wide range of modeling errors with respect to the underlying manipulation task.
APPROACH SURROUND TRANSPORT Near_Object ~Closure ! ~Quorum Closure " Quorum ~Near_Object
Figure 3.1: Behavior architecture. The software for each robot is identical and consists of several simple modes and the sequential composition of these modes.