A topic not yet mentioned in this thesis is the interoperability of robots from different projects that try to achieve similar goals. This research area lies beyond the scope of this thesis. Nevertheless, Miro was already applied for conducting early experiments in this new and upcoming scientific topic as first discussed in [141].
Due to scientific as well as pragmatic reasons, there is a growing interest in the robotics field to join the efforts of different labs to form mixed teams of autonomous mobile robots. In RoboCup, the pragmatic reasons are com- pelling. The recent rule change in the mid-size league allows for more robots per team, and in the RoboCup Rescue league a group of heterogeneous robots with diverse capabilities is likely to perform better than one system that tries to encapsulate them all. However, the limited financial resources and the ad- ditional maintenance effort for further robots exceeds the capabilities of many research labs. Also, the threshold for new research groups to participate in long term scenarios such as RoboCup is lowered if they only need to contribute one or two robots to a mixed team, instead of having to build an entire team. Mixed teams are also motivated from a scientific perspective. They introduce the research challenge of cooperation within teams of extremely heterogeneous autonomous mobile systems.
As most robots in the middle-side league are custom built, or at least customized commercial research platforms with unique configurations of actuator and sen- sor configurations, mixed teams from different laboratories are extremely het- erogeneous. There are few commonly used high-level libraries for sensor data processing and reactive actuator design in the community. Furthermore there is a multitude of methods and schools, each deliberately designing the con- trol architecture of their robots fundamentally different to their competitors. This makes the unification of the software of the different robots of a potential
9.6. MULTIROBOT TEAMS SPANNING MULTIPLE LABS 117
Ulm Munich Graz
Figure 9.3: Individual robot platforms from the different labs.
mixed team almost impossible without substantial rewriting of at least one of the team’s software. In our opinion it would be also undesirable. Why should an autonomous mobile robot have to commit to any kind of sensor processing or control paradigm to be able to cooperate with another team mates, if both are programmed to interact in the same problem domain?
The basic idea of the conducted experiments in mixed robot teams was, that for cooperation between robots, the sharing of information about the environ- ment is initially sufficient for successful cooperation. If all robots share both the same belief about their environment, as well as the same set of goals, simi- lar conclusions should be drawn. This is known as the Intentional Stance [22], and has proven to be a successful way of coordinating behavior in RoboCup scenarios [130, 18]. A central prerequisite for successful team cooperation was therefore the unification of the beliefs about the world of the different agents. The limitations of the individual sensors usually provide each robot with quite limited information about the state of its environment. So it is unlikely that the beliefs derived solely from the robots’ own sensors are automatically suf- ficiently similar to coordinate behavior in a shared environment. Sharing of information was therefore considered essential for solving this problem. As dif- ferent autonomous mobile platforms robots are equipped with different sensor suites that each provide their own unique perception of the environment, this improves the quality of the information available to the robots, even if they ob- serve exactly the same scene from similar positions. Compare for instance laser range finders, which provide precise depth information, with color cameras, which provide more certainty about object identity.
At the RoboCup WorldCup 2004 in Lisbon a heterogeneous mixed-team exper- iment was conducted. For this purpose, one robot of the The Ulm Sparrows joined the teams of the middle-size league teams of the TU Munich, the “Agilo Robocuppers” [118] and the TU Graz team, “Mostly Harmless” (see Figure 9.3). For Agilo, Ulm provided the goal-keeper, in the Graz team, a striker was added. All teams used the group communication facility provided by Miro to exchange the beliefs of each robot with the other team mates. For this purpose,
118 CHAPTER 9. RESULTS AND APPLICATIONS
GoalPost Ball Robot
Dynamic Object Teamate 1..N Opponent 1..N Opponent Teammate GoalPostOppL/R GoalPostOwnL/R GoalPostOwn GoalPostOpp CornerFlagOppL/R CornerFlagOwnL/R CornerFlagOpp CornerFlagOwn CornerFlag Static Object Object
Figure 9.4: Tree of object classes for robot soccer.
a common representation for the robots observations was designed (the so-called belief state). It defines a classification hierarchy on the objects present in the robot soccer world (Figure 9.4), and models uncertainty as well as precision of the observed objects. In a followup to this experiments, a diploma thesis at the TU Munich currently evaluates methods of learning team-coordination based solely on a fused belief state and knowledge about the individual properties of the other robots in the team.