Margaret E. Jefferies

Computing a representation of the local environment

Yeap (1988) argued that an important basis for computing a cognitive map is the ability to compute and recognise local environments. Although he has demonstrated how such local environments could be used to construct a raw cognitive map, he failed to produce an adequate algorithm for computing them. In this paper, a detailed study of this problem is presented. We argue that although each local environment computed forms a natural basis for constructing a raw cognitive map, it is not computed primarily to do so. Instead, it is computed for ones immediate needs (such as hunting a prey or escaping from danger). This change in perspective argues for a very different cognitive mapping process, namely one that computes local environments as the individual moves through the environment but these representations are not necessarily used to construct a raw map. The individual does not do so until there is evidence that it is going to stay. Consequently this simplifies the algorithm for computing a local environment and a new algorithm is thus proposed. Some results of our implementation are shown.

Robotics and Cognitive Approaches to Spatial Mapping

This important work is an attempt to synthesize two areas that need to be treated in tandem. The book brings together the fields of robot spatial mapping and cognitive spatial mapping, which share some common core problems. One would expect some cross-fertilization of research between the two areas to have occurred, yet this has begun only recently. There are now signs that some synthesis is happening, so this work is a timely one for students and engineers in robotics.

The Utility of Global Representations in a Cognitive Map

In this paper we propose the use of small global memory for a viewers immediate surroundings to assist in recognising places that have been visited previously. We call this global memory a Memory for the Immediate Surroundings (MFIS). Our previous work [1, 2] on building a cognitive map has focused on computing a representation for the different local spaces the viewer visits. The different local spaces which are computed can be connected together in the way they are experienced to form a topological network which is one aspect of a cognitive map of the spatial environment. The problem with topological representations is that using them one cannot easily detect that one is reentering a previously visited part of the environment if it is approached from a different side to the one used previously. Thus we have developed a cognitive map representation which comprises an MFIS working in cooperation with the topological network. The idea that a global map is present as part of the cognitive mapping process is increasingly appealing. Robotics researchers have used them from the early days of autonomous mobile robots. However, they have shown that it is difficult to compute an accurate global representation because of errors. There is now increasing evidence that a global map is used in animals and many simulation models have incorporated the use of such a map. In this paper we review these works, discuss this notion of a global map in cognitive mapping, and show how one could be computed with minimum effort.

Robot Cognitive Mapping – A Role for a Global Metric Map in a Cognitive Mapping Process

In robotics it would be argued that we are closing the loop in a topological map using a global metric map. Drawing on our studies of human and animal cognitive mapping we proposed that a cognitive map comprises a topological map of metric local space representations [24]. Each local space defines a part of the environment that appears to enclose the animal/robot. Recently our Pioneer 2 robot has been computing such a map during its travels around our department. The advantage of such a map for a robot is that cumulative positional error is constrained to the local representation. Simpler localisation methods will often suffice for the local environment as global metric consistency is not required. The trade-off is that one cannot easily detect that one is re-entering a previously visited part of the environment via a new route (i.e. closing a loop) as is the case with a global metric map. The question we asked was: could we combine the local and global representations, exploiting the advantages of both - local representations for simpler localisation and no global metric consistency; global representation for easy loop detection. While a simple localisation method suffices for the local representation it would be inadequate for a global metric map. However the error could not grow unbounded if it were to be useful in the task of detecting loops. Our solution was to limit the size of the global map and have it move with the robot as it traversed its environment. We will describe the implementation of such a map and show that it can detect loops over a reasonable distance.

Using 2d and 3d landmarks to solve the correspondence problem in cognitive robot mapping

We present an approach which uses 2D and 3D landmarks for solving the correspondence problem in Simultaneous Localisation and Mapping (SLAM) in cognitive robot mapping. The nodes in the topological map are a representation for each local space the robot visits. The 2D approach is feature based – a neural network algorithm is used to learn a landmark signature from a set of features extracted from each local space representation. Newly encountered local spaces are classified by the neural network as to how well they match the signatures of the nodes in the topological network. The 3D landmarks are computed from camera views of the local space. Using multiple 2D views, identified landmarks are projected, with their correct location and orientation into 3D world space by scene reconstruction. As the robot moves around the local space, extracted landmarks are integrated into the ASRs scene representation which comprises the 3D landmarks. The landmarks for an ASR scene are compared against the landmark scenes for previously constructed ASRs to determine when the robot is revisiting a place it has been to before.

Using context to solve the correspondence problem in simultaneous localisation and mapping

We present a method for solving the correspondence problem in Simultaneous Localisation and Mapping (SLAM) in a topological map. The nodes in the topological map are a representation for each local space the robot visits. The approach is feature based - a neural network algorithm is used to learn a signature from a set of features extracted from each local space representation. Newly encountered local spaces are classified by the neural network as to how well they match the signatures of the nodes in the topological network. Of equal importance as the correspondence problem is its dual, that of perceptual aliasing which occurs when parts of the environment which appear the same are in fact different. It manifests itself as false positive matches from the neural network classification. Our approach to solving this aspect of the problem is to use the context provide by nodes in the neighbourhood of the (mis)matched node. When neural network classification indicates a correspondence then subsequent local spaces the robot visits should also match nodes in the topological map where appropriate.

The Correspondence Problem in Topological Metric Mapping – Using Absolute Metric Maps to Close Cycles

In Simultaneous Localisation and Mapping (SLAM) the correspondence problem, specifically detecting cycles, is one of the most difficult challenges for an autonomous mobile robot. In this paper we show how significant cycles in a topological map can be identified with a companion absolute global metric map. A tight coupling of the basic unit of representation in the two maps is the key to the method. Each local space visited is represented, with its own frame of reference, as a node in the topological map. In the global absolute metric map these local space representations from the topological map are described within a single global frame of reference. The method exploits the overlap which occurs when duplicate representations are computed from different vantage points for the same local space. The representations need not be exactly aligned and can thus tolerate a limited amount of accumulated error. We show how false positive overlaps which are the result of a misaligned map, can be discounted.

Using absolute metric maps to close cycles in a topological map

In simultaneous localisation and mapping (SLAM) the correspondence problem, specifically detecting cycles, is one of the most difficult challenges for an autonomous mobile robot. In this paper we show how significant cycles in a topological map can be identified with a companion absolute global metric map. A tight coupling of the basic unit of representation in the two maps is the key to the method. Each local space visited is represented, with its own frame of reference, as a node in the topological map. In the global absolute metric map these local space representations from the topological map are described within a single global frame of reference. The method exploits the overlap which occurs when duplicate representations are computed from different vantage points for the same local space. The representations need not be exactly aligned and can thus tolerate a limited amount of accumulated error. We show how false positive overlaps which are the result of a misaligned map, can be discounted.

Neural network approaches to cognitive mapping

There are many different approaches to cognitive mapping, arising mostly from the different philosophical backgrounds of the researchers involved. Our own research into the problem of how best to build a representation for ones experience of ones spatial environment is motivated by the need to understand how the human mind works. Neural network approaches to cognitive mapping are as varied as their non-neural network counterparts and range from models which use the network to model the physiology of the brain to models which are merely an abstraction of some aspect of cognitive mapping behaviour. We review four neural network approaches to cognitive mapping with the view to determining what insights they can bring to the cognitive mapping process.

Computing the local space of a mobile robot

A popular approach to describing the environment of an autonomous system is to compute a representation for the space surrounding the robot, termed the local space. Recently the focus of much of the work in this area in robotics has been on acquiring a usable representation. To this end many computationally demanding algorithms have been devised in the hope that accurate representations which more closely match the real world will be computed. However this is very difficult to achieve from the robot’s initial experience of its environment. We argue that an inaccurate but useful representation can be computed from the robot’s initial view of the local space. We present an algorithm for computing this initial representation and show its implementation on a robot with sonar sensors.