In this project, we proposed Action-Related Place (ARPlace) as a powerful and flexible representation of task-related place in the context of mobile manipulation. ARPlace represents base places not as a single position, but rather as a collection of positions, each with an associated probability that the manipulation action will succeed when standing there. ARPlaces are generated using a predictive model that is acquired through experience-based learning. As this model is learned from observed successful and failed grasps, it is grounded in the robot's actual behavior. We have integrated ARPlaces in a symbolic transformational planner, and our evaluations demonstrate that using ARPlaces leads to more robust and efficient mobile manipulation in the face of state estimation uncertainty.

Optimizing the Execution of Symbolic Robot Plans

The use of abstract symbolic actions enables robots to employ small action libraries, which reduces the search space for decision making. When executing the actions, however, the robot must tailor the abstract actions to the specific task and situation context at hand. In this project we proposed a novel robot action execution system that learns success and performance models for possible specializations of abstract actions. At execution time, the robot uses these models to optimize the execution of abstract actions to the respective task contexts. The robot can so use abstract actions for efficient reasoning, without compromising the performance of action execution. We show the impact of our action execution model in three robotic domains and on two kinds of action execution problems: (1) the instantiation of free action parameters to optimize the expected performance of action sequences; (2) the automatic introduction of additional subgoals to make action sequences more reliable.

Main Cooperation Partners: Michael Beetz (PhD. Supervisor) Institute: Technische Universität München Duration: 2003-2008 Further Infos: Related Publications Implicit Coordination in Teams of Robots A striking aspect of human coordination is that we achieve it with little or no communication. We achieve this implicit coordination by taking the perspective of others, and inferring their intentions. In contrast, robots usually coordinate explicitly through the extensive communication of utilities or intentions. In this project we developed a method that combines both approaches: implicit coordination with shared belief. In this approach, robots first communicate their beliefs about the world state to each other, using a CORBA based communication module. They then use learned utility prediction models to predict the utility of each robot locally. Based on these utilities, an action is chosen. Within a heterogeneous soccer team, with robots from both the Munich and Ulm RoboCup mid-size teams, we applied implicit coordination with shared belief to a typical task from robotic soccer: regaining ball possession. An empirical evaluation demonstrates that the redundancy of implicit coordination with shared belief leads to robustness against communication failure and state estimation inaccuracy. Main Cooperation Partners: Michael Isik, Hans Utz, Gerd Maier Institute: Technische Universität München University of Ulm Duration: 2003-2007 Further Infos: Related Publications, Movie Learning Objective Functions for Face Model Fitting Model-based fitting has proven to be a successful approach to interpreting the large amount of information contained in images. Fitting algorithms search for the global optimum of an objective function, which should correspond to the best model fit in a given image. Although fitting algorithms have been the subject of intensive research and evaluation, the objective function is usually designed ad hoc, based on implicit and domain-dependent knowledge. This often leads to functions with many local minima, and a global minimum that does not correspond to the best model fit. We address the root of this problem by learning more robust objective functions. First, we formulate a set of desirable properties for objective functions, and give a concrete example of an ideal function that has these properties. Then, we propose a novel approach that learns an objective function from training data generated by this function and manually annotated images. In this approach, critical decisions such as the feature selection are automated, and the remaining manual steps hardly require domain-dependent knowledge. An extensive empirical evaluation demonstrates that learned objective functions enable fitting algorithms to determine the best model fit more accurately and efficiently than designed objective functions. Main Cooperation Partners: Matthias Wimmer, Christoph Mayer, Bernd Radig Institute: Technische Universität München Duration: 2006-2008 Further Infos: Related Publications Completion of Occluded Surfaces in Range Images Analysis and reconstruction of range images usually focuses on complex objects completely contained in the field of view; little attention has been devoted so far to the reconstruction of partially occluded simple-shaped wide areas like parts of a wall hidden behind furniture pieces in an indoor range image. The work in this project is aimed at such reconstruction. First of all the range image is partitioned and surfaces are fitted to these partitions. A further step locates possibly occluded areas, while a final step determines which areas are actually occluded. The reconstruction of data occurs in this last step. Main Cooperation Partners: Bob Fisher, José Santos-Victor Institute: University of Edinburgh Instituto Superior Tecnico Duration: 09.2000-12.2001 Funding: EU TMR Network CAMERA Further Infos: Related Publications Knowledge-based Analysis of the Transmission Control Protocol Using a knowledge-based approach, we derive a protocol for the sequence transmission problem, which provides a high-level model of the Transmission Control Protocol, which is primarily used in Internet communication. The knowledge-based protocol is correct for communication media where deletion and reordering errors may occur. Furthermore, it is shown that both sender and receiver eventually attain depth N knowledge about the values of the messages for any N, but that common knowledge about the messages is not attainable. Main Cooperation Partners: Rineke Verbrugge Institute: University of Groningen Duration: 2000-2002 Further Infos: Related Publications, Applet