General Concepts for Human Supervision of Autonomous Robot Teams
For many dangerous, dirty or dull tasks like in search and rescue missions, deployment of autonomous teams of robots can be beneficial due to several reasons. First, robots can replace humans in the workspace. Second, autonomous robots reduce the workload of a human compared to teleoperated robots,...
| Main Author: | |
|---|---|
| Format: | Others |
| Language: | German en |
| Published: |
2014
|
| Online Access: | http://tuprints.ulb.tu-darmstadt.de/3873/7/dissertation.pdf Petersen, Karen <http://tuprints.ulb.tu-darmstadt.de/view/person/Petersen=3AKaren=3A=3A.html> : General Concepts for Human Supervision of Autonomous Robot Teams. Technische Universität, Darmstadt [Ph.D. Thesis], (2014) |
| Summary: | For many dangerous, dirty or dull tasks like in search and rescue missions, deployment of autonomous teams of robots can be beneficial due to several reasons. First, robots can replace humans in the workspace. Second, autonomous robots reduce the workload of a human compared to teleoperated robots, and therefore multiple robots can in principle be supervised by a single human. Third, teams of robots allow distributed operation in time and space. This thesis investigates concepts of how to efficiently enable a human to supervise and support an autonomous robot team, as common concepts for teleoperation of robots do not apply because of the high mental workload. The goal is to find a way in between the two extremes of full autonomy and pure teleoperation, by allowing to adapt the robots’ level of autonomy to the current situation and the needs of the human supervisor. The methods presented in this thesis make use of the complementary strengths of humans and robots, by letting the robots do what they are good at, while the human should support the robots in situations that correspond to the human strengths.
To enable this type of collaboration between a human and a robot team, the human needs to have an adequate knowledge about the current state of the robots, the environment, and the mission. For this purpose, the concept of situation overview (SO) has been developed in this thesis, which is composed of the two components robot SO and mission SO. Robot SO includes information about the state and activities of each single robot in the team, while mission SO deals
with the progress of the mission and the cooperation between the robots. For obtaining SO a new event-based communication concept is presented in this thesis, that allows the robots to aggregate
information into discrete events using methods from complex event processing. The quality and quantity of the events that are actually sent to the supervisor can be adapted during runtime by defining positive and negative policies for (not) sending events that fulfill specific criteria. This reduces the required communication bandwidth compared to sending all available data.
Based on SO, the supervisor is enabled to efficiently interact with the robot team. Interactions can be initiated either by the human or by the robots. The developed concept for robot-initiated interactions is based on queries, that allow the robots to transfer decisions to another process or the supervisor. Various modes for answering the queries, ranging from fully autonomous to pure human decisions, allow to adapt the robots’ level of autonomy during runtime. Human-initiated interactions are limited to high-level commands, whereas interactions on the action level (e. g., teleoperation) are avoided, to account for the specific strengths of humans and robots. These commands can in principle be applied to quite general classes of task allocation methods for autonomous robot teams, e. g., in terms of specific restrictions, which are introduced into the system as constraints. In that way, the desired allocations emerge implicitly because of the introduced constraints, and the task allocation method does not need to be aware of the human supervisor in the loop. This method is applicable to different task allocation approaches, e. g.,
instantaneous or time-extended task assignments, and centralized or distributed algorithms.
The presented methods are evaluated by a number of different experiments with physical and simulated scenarios from urban search and rescue as well as robot soccer, and during robot competitions. The results show that with these methods a human supervisor can significantly improve the robot team performance. |
|---|