WP3: Qualitative spatial cognition

Spatial models of the environment are at the very core for mobile robot systems. Robots are currently in the progress of moving out from the factories and into our homes and offices to, for example, run errands for us or otherwise assist us. Communication between robots and humans will therefore become increasingly important.

To support this the robot does not only have to be able to perform navigation, but must go beyond that. It has to have an understanding of the environment that allows it to answer questions such as, for example, “what functiondoes this part of the space fulfil?”, “where would I typically find object X?”, “what can be done with object X?”. That means that, on top of a quantitative model of space, the robot needs to have a conceptual model of spatial entities – rooms or other topological units, objects, their individualproperties, and their relations. If the robot is to perform its tasks in a man-made, human-populated environment, and, moreover, if the robot is to communicate about its environment with humans, then the conceptual spatial model needs to reflect how humans conceptualise their environment.

Comparing how robots and humans perceive and represent the world shows that there is a large
difference. For communication between the two to work, this gap has to be bridged. Robots have to
start perceiving things closer to the way humans do, especially when interacting with real end-users
and not the scientists that designed the robot system.

Our hypothesis is that objects play an important role when building a spatial model of a man-made environment for interaction with humans and answering questions like the ones above. Spatial models with objects as core building blocks would allow for a scalable
representation and act as the basis for much of the high level reasoning. The spatial modelling and the search for objects are naturally coupled. When looking for a certain object, the search can be directed towards areas in the environment where such objects are normally found. For example, when looking for a box of cereals, the floor is an unlikely place to find it.

Furthermore, since manipulation of object is of special interest in this project we will primarily work with objects that can be manipulated by the system and focus on objects located within graspable distance of the system. Another important question in conjunction with the spatial modelling is
the question of short-term versus long-term memory. Traditional robot systems often rely on a
global long-term representation for the environment, the map, that allows the robot to stay localised
and plan its way from one part of the environment to another. For local navigation another local
but much more detailed map is often used. This local map acts as a detailed short-term memory that
the robot will forget as soon as this particular area is left. The global model of the environment can be layered to support different functions at different levels of abstraction, from low level navigation to conceptual reasoning.
The main objectives of this WP are:

• Study how to best incorporate objects into the spatial model of the environment?
• Investigate how the object-based spatial models can be used to infer knowledge about type and function of an area, typical placement of object classes, etc?
• Study how to perform spatial referencing between object in the model and relative to the robot?
• Find out what part of the environment should be captured in short-term
representations and what goes into long-term memory and investigate if the same layering is
necessary for the short-term memory as for the long-term memory?
• Study how to utilise and learn spatial relations in human-robot interaction?
• Investigate how to represent gaps in spatial knowledge?