Differences

This shows you the differences between two versions of the page.

Link to this comparison view

Next revision		Previous revision
object_pose_representation [2012/12/01 14:23] – created tenorthobject_pose_representation [2014/06/05 11:38] (current) – external edit 127.0.0.1
Line 1: Line 1:
-Information about the poses and dimensions of objects is crucial for finding and manipulating +#REDIRECT doc:object_pose_representation
-them. In K NOW ROB, object dimensions are described as simple bounding boxes or cylinders +
-(specifying the height, and either width and depth or the radius). While this is clearly not suf- +
-ficient for grasping, we chose this description as a compromise in order not to put too many +
-details like point clouds or meshes into the knowledge base. Such information is rather linked +
-and stored in specialized file formats. +
-Object poses are described via homography matrices. Per default, the system assumes all +
-poses to be in the same global coordinate system. Pose matrices can, however, be qualified with +
-a coordinate frame identifier. The robot can then transform these local poses into the global +
-coordinate system, for example using the tf library2 . +
-Since robots act in dynamic environments, they need to be able to represent both the current +
-world state and past beliefs. A naive approach for describing the pose of an object would be +
-to add a property location that links the object instance to a point in space or, more general, a +
-homography pose matrix. However, this approach is limited to describing the current state of the +
-world – one can express neither changes in the object locations over time nor differences between +
-the perceived and an intended world state. This is a strong limitation: Robots would not be able +
-to describe past and (predicted) future states, nor could they reason about the effects of actions. +
-Memory, prediction, and planning, however, are central components of intelligent systems. +
-The reason why the naive approach does not support such qualified statements is the limitation +
-of OWL to binary relations that link exactly two entities. These relations can only express if +
-something is related or not, but cannot qualify these statements by saying that a relation held +
-an hour ago, or is supposed to hold with a certain probability. For this purpose, we need an +
-additional instance in between that links e.g. the object, the location, the time, and the probability. +
-In K NOW ROB, these elements are linked by the event that created the respective belief: the +
-perception of an object, an inference process, or the prediction of future states based on projec- +
-tion or simulation. The relation is thus reified, that is, transformed into a first-class object. These +
-reified perceptions or inference results are described as instances of subclasses of MentalEvent +
-(Figure 3.4), for instance VisualPerception or Reasoning. Object recognition algorithms, for in- +
-stance, are described as sub-classes in the VisualPerception tree. Multiple events can be assigned +
-to one object, describing different detections over time or differences between the current world +
-state and the state to be achieved (Figure 3.5). +
- +
-{{ :mental-events.png?nolink&300 |}} +
- +
-{{ :internal-object-representation.png?nolink&300 |}} +
- +
-This representation is similar to the fluent calculus [Thielscher, 1998], in which fluents are +
-objects that represent the change of values over time. In our case, however, the reified objects +
-contain more information than just a changing value: the current and all past states of the relation, +
-including the times at which state changes were detected, and the type of event that established +
-the relation. Using our representation, we can describe multiple “possible worlds”, for example +
-the perceived world, a description of how the world is supposed to look like, and the world state +
-a robot predicts as the result of some actions it performs. Since all states are represented in the +
-same system, it becomes possible to compare them, to check for inconsistencies or to derive +
-the required actions, which would be difficult if separate knowledge bases would be used for +
-perceived and inferred world states. +
- +