Advantages of Qualitative Reasoning to Robotics

Current robot planning systems operate on a precise geometric and numeric description of the robot, its workspace, and the objects that it is manipulating. Programmable robots (as opposed to guided, or ``lead-through'' robots) are programmed to perform motions according to this precise description, which seldom includes any tolerance information. Adequate precision of the robot in carrying out its program, and exact location of objects in the workspace is therefore a major concern of industrial robotics, because sufficient precision must be obtained so that the numeric description agrees with the workspace.

Despite impressive achievements of precision in robots, there are many problems which cannot be solved by simply increasing precision. It is these problems which can benefit from the use of qualitative spatial reasoning methods instead of numerical geometry. A few important problems for advanced robot programming that may be solvable by the application of qualitative methods are as follows:

• A robot will encounter errors in its data from time to time, regardless of normal operating precision. In these situations it should be able to continue operating, perhaps less efficiently, rather than having to abandon the task. This is known as ``graceful degradation'' in performance.
• High level robot programming should be carried out without having to refer to numeric data. Ideally, a robot programmer should describe the task to the robot in terms that they would use to describe it when doing it themselves (``task-level'' programming). People do not naturally think of physical actions in terms of joint angles or numeric workspace co-ordinates, so high level robot programming should be done in non-numeric terms.
• A common way for people to communicate information about spatial tasks is by the use of diagrams. A diagram normally reflects the structure of the task only, without containing any important dimensions in the actual lines of the diagram - it is a qualitative device. A robot should be able to understand the structure of the task from this kind of information.
• Where parts of a robot's workpiece are hidden, and the dimensions are therefore completely unknown, the robot must be able to make hypotheses about their shape. Consider the removal of a key from a keyhole, for instance - we must make some mental picture of what the hidden part of the key might look like, before developing a strategy for removing it. If numerical information is needed for robot operations, an hypothesis must include a complete geometric description, whereas a qualitative reasoning system allows the robot to proceed on simple structural hypotheses.
• If the task involves design, it is often necessary to propose hypothetical values for some design parameters (where the design is underconstrained). In this case, the hypothetical assumptions made should include no more information than is absolutely necessary to continue design, since added information may cause the hypothesis to fail unnecessarily. The use of a qualitative representation can provide this facility, by describing only essential structural features.

An example of a current project in high level robot reasoning which could benefit from the use of qualitative representations is Andreae's NODDY system [And85], which forms the basis of an ongoing project at Victoria University. This system observes the actions of a conceptual robot in a geometric world, and learns about robot procedures by generalising from its observations. Numeric information is largely irrelevant to acquiring the functional aspects of useful procedures, although NODDY presently uses a number of numerical techniques. A qualitative description of robot actions would provide a set of information for generalisation to proceed from that had already been filtered to isolate structural elements.

There are a wide range of robot reasoning tasks which are hampered by the complexity of operations in three dimensional numerical geometry. The original starting point of my research was a proposal to investigate the understanding of fasteners through robot disassembly of real mechanical devices. The proposal excluded the physical issues of vision and manipulation, but still resulted in a list of almost 100 research areas which would need substantial progress before a start could be made on the complete disassembly problem.^1.1

Many of these disassembly and fastening research issues derived their complexity from the fact that representations available in three dimensional robotic reasoning systems were not appropriate to the task that I wished to solve. This failing originated the investigation of qualitative spatial reasoning that is described in this thesis.