Sensory Systems/Insects/Navigation

Spatial Memory in Insect Navigation
Some insects show a remarkable memory of location. Crickets learn where a cool spot is located in an otherwise hot area. Parasitoid wasps (Argochrysis armilla) can remember where another hymenoptera (Ammophila pubescens) dug its inconspicuous nest in order to feed on its larvae. To help other bees find their way back into the nest Partamona batesi has developed an elaborate behaviour which consists of sticking white river sand together to form a good visible portico. All these different behaviours have in common that the insects have a sense of their environment and can navigate within it. Spatial memory is observed in the behaviour of various insects with central place foraging. Remembering locations is essential for finding and remembering resourceful food sites, returning to nests, or to hold position in flowing water or in the air [5].

Until recently behavioural experiments were mainly performed with social bees and ants, and studies in simplified environments (with reduced amounts of landmarks and altered panoramas depending on the question asked) to unravel three different types of memory-based guidance mechanisms [5]. While two are based on memory of views of the surrounding, the third is based on an inner accumulator creating a vector from the nest to a desired site [2][3][4][5][7][12][13]:


 * Alignment image-matching is used to head along familiar routes, comparing memorised snap-shot views with the present sight [5][10].


 * Position image-matching is a more general way of orientation and is used when the desired goal is known but the starting position or the route is new [1][5][13].


 * Path integration is exerted when the perceived environment is unknown or when travelling through featureless landscapes. The insect measures the distance and the direction creating a vector [2][5].

Using all three procedures, the insect compares its sensory input with a memory of the desired sensory input. The discrepancy of these inputs is transformed into an “output vector” giving a direction towards the desired goal [4][5]. Depending on the situation not all mechanisms are needed simultaneously. Orientation over a range of conditions is achieved by converging and complementing of the three different processes [4][5].

Alignment Image-Matching
Alignment image-matching is the most basic way to compare visual input to memory. It is used when travelling along a known route. Doing so, a current retinal picture is compared to visual memories and made congruent [10].

Fractional Position of Mass, Orientated Edges and Segmentation
Lent et al. looked at the behaviour of wood ants (Formica rufa) in an artificial scene, comparing the routes taken in the original and in the altered scene. The group found several different behaviours which helped the ant choose a direction. One of these processes resembles an already known ability of ants to compute the centre of mass of shapes. It was found that ants orientate themselves using the “fractional position of mass” (FPM) which gives a robust orientation over several meters distance from the shape. The FPM is a direction to which the ant heads to and is described by a ratio of the left to the right area of the shape (Figure 1). Further it was found that the insects orientate themselves by extracting local visual features such as oriented edges (for example a diagonal margin of a shape) and superimpose them on a visual memory [8][10]. Additionally the ants seem to segment complex scenes calculating the FMP of each piece individually. In simple scenes one FPM is calculated over the whole image [10].

For optimal guidance it was found that the ant orientates itself first by local features. A recent study suggets that ants probably segment their view and calculate the corresponding FPM. If the insects diverge from their original direction, a correction of the body orientation is achieved by saccade-like turns [10]. It has been experimentally assessed that wood ants do this alignment every three seconds, correcting up to 70 degrees in direction if necessary [5].

Landmarks and Skylines
The characteristics of a reliable landmark are the availability over several journeys and its visibility in various different light conditions (Figure 2). Graham and Cheng have suggested that the skyline is such a trustable landmark. A skyline profile is an object on the ground which contrasts against the sky [7]. Their experiments were conducted using so-called zero vector ants. Zero vector ants are ants caught right before entering their nest. Due to this they lack a path integration vector (which is set at zero) and have to navigate with visual cues only when displaced. Graham and Chen investigated how the ants successfully home on their nest when displaced. Additionally this behaviour was also tested in an altered environment with an artificial panorama, and with several sideways shifted versions of it. They found that ants shifted their direction according to the false skyline independent of an inner compass mechanism [7].



Other researchers differentiate the omnipresent skyline from other landmarks. Wystrach et al. observed that the guidance of ants (Melophorus bogati) cannot be completely explained by orientating with landmarks only. Since an ants' eye has a poor resolution and thus reduces complex natural sceneries, it has been suggested that panoramas or skyline views are used for basic orientation, giving primary crude directional hints [12].

Positional Image-Matching
Positional image-matching describes a more universal utilisation of visual memories. Compared to alignment image-matching this process allows guidance from new, unknown locations via novel directions to a desired goal as long as there are enough familiar elements present [5]. In cluttered environments ants rely heavily on their path integrator vector and on surrounding landmarks. If they are in familiar terrain with sufficient landmarks, the path integrator vector is neglected and the insects rely on the snapshots [1][13]. That this process might be used by ants was shown in simulations in a robot performing image-matching which showed similar behaviour to ants[11].

Skyline Heights, Visual Compass and Mismatch Gradient Descent
Wystrach et al. have proposed three further view-matching strategies. Firstly they suggest that ants compare the heights of the skylines of the present location with the memory of the view at the goal scene. If the skyline is too high at the current state, then the ant has to walk in the direction it is looking. In contrast, if the skyline is too low, the ant has to veer away from the current direction of orientation (Figure 3). Notwithstanding the general robustness of this model, it has two flaws. First, the height difference between the two views must be big enough to tell a difference. Second, the ant has to have the same absolute orientation as in the memory view in order to successfully compare the two skylines. This however may be achieved via a geomagnetic or a celestial compass. Therefore skyline height comparison is a good tool to obtain the rough heading direction when in a new location, but it is barely practical when near the goal [13].



Second, they proposed that ants have a visual compass which requires memories of several different views taken around the nest while facing it. If the insect finds itself in an unknown position, it can compare the current view with the memorised views and pick the best matching one and calculate the heading direction out of that. This model is supported by the observation that Ocymymex ants perform a well-choreographed learning walk which also involves facing the nest. Compared to the skyline height comparison, ants do not have a problem finding the appropriate orientation, because their memory views are already directed to the nest. However if the retained memory view of the nest is not the best matching, the ant may be led in a completely wrong direction. This means that using the visual compass on a familiar route is a very robust model, but not when it comes to finding a way to the nest from an unfamiliar location [13].

The third proposed process is called mismatch gradient descent and provides orientation. The perceived and the memorised images are constantly compared to each other. In order for this process to work, ants would need three dimensional information which would have to be obtained by moving diagonally through the landscape. However this has not been observed and leads to the conclusion that mismatch gradient descent alone would not render the ant at the wanted destination [13].

These findings suggest that the skyline height comparison should be used to find a heading direction. Mismatch gradient descent helps the ant remember its direction to the nest during walking and once near the nest, the visual compass leads the ant to its homestead. All the different models can be mixed, their relative contribution probably varying [13]. If none of the mechanisms proposed above indicate a heading direction, i.e. if the insect is in a completely new environment, ants start with a systematic search consisting of loops which increase in size the more time passes during the search [13].

Confusion After Displacement
It was shown that zero vector ants which are trained along a route exhibit a short phase of confusion expressed by unorientated walking after they are picked up in front of their home and replaced at the feeder. It seems as if the route memory towards the nest was ignored during the time of disorientation. It is evident that the repetition of the route caused the confusion of the animal letting the question arise, whether they have an episode-like memory which tells them that they already have taken the route [4]. However, the confusion is not always observed, especially not in cluttered environments. This could be explained by the assumption that ants segregate their view in cluttered environments in stages. When moved, ants may think that they are still on their way to the nest since the last stage of the route has not been reached yet [4].

Path Integration
Path integration is a kind of dead-reckoning where there is ongoing localisation via measurement of the direction, the velocity and the time taken [2][5]. This requires an accumulator as a core component of the whole system [6]. As a result the ant obtains a vector with the resulting distance and direction it went relative to the nest's position [5]. To know headings for a desired site, an output vector is generated by subtracting the current path integration state from the memory of the path integration state at the goal [4][5].

Build Up of an Path Integrator Vector
In order to build a vector the insect needs a starting point. The convention of this model is that the nest is taken as the point of origin and defined as zero [6].The direction of the vector has to be made in relation to external or internal cues. External cues are the sun compass, large landmarks or polarised light [3]. Internal cues such as a sense for the magnetic field of the earth are used if there is nothing in the environment to navigate [2][3][5]. By bees, the length of the vector is identified by measuring the distance travelled using the optic flow. Ants do this by the proprioceptive input derived from their steps [2][5].

Ants Completely Rely on Path Integration in Featureless Environments
Because of the arithmetic way to orientate, insects using path integration can move in completely unfamiliar terrain. This helps the animals to travel through featureless environment and helps them return to their nest after an exploration journey [5].

A simple experiment to demonstrate that ants completely rely on path integration was shown by Collett and Collett. Returning ants were displaced into either familiar or unknown areas before they reach their goal. In both cases the ants continue their route until they reach the point where they should have reached the nest. Only then do they start to show search behaviour [6].



Path Integration Is a Mathematical and Cognitive Challenge
Next to an accumulator which keeps track of the current position, insects using path integration need some mechanism to store path integrator states of desired places (for example at feeding grounds) for subsequent remembrance. In total, four factors need to be remembered during navigation: the present state of the path integrator, the path integration state of the goal, the output of a comparator and the output vector. While this kind of guidance system is clearly based on vector computation, no neural circuits are known which could process this [6].

Two Main Models for Path Integration
There are two main models known for path integration. The first postulates that the accumulator gets updated continuously, while the second suggests an intermitted accumulator update. In the model of the continuous updated accumulator, the accumulator state upon returning has to be the same as when leaving the nest. Therefore the coordinate system is in-bound. In the intermitted updating model, the accumulator state gets reset at the nest and at the point where the ant turns around to return. To return, the polarity of the compass gets inverted and the ant heads in the direction of the opposite vector, therefore the system is out-bound. Studies suggest that the latter model is more probable since the outward and inward vector do not exactly have to be cancelled out leaving room for small errors [6].

Adjustment of Path Integration and Image-Matching
As discussed above, insects do not only rely on path integration during navigation. Image-matching based on landmarks is the major way of orientation on familiar routes. Path integration and image matching work independently but can also interact, adjusting each other. During the course of walking directed by path integration, new landmarks can be learned. And during walking with image-matching the navigation vector can be standardized [6].

Path Integration and Research
Insect navigation is also very interesting for robot developers. Several Robots have already been constructed using path integration as one of their main navigational systems. It is also a principle which is applied by driverless cars [2][3][11]. Lambrinos et al. constructed a robot called “Sahabot” which orientates itself based on path integration and view based matching [9]. Navigation in wide open areas seems to be less complicated than in cluttered environments, which is still subject to intense research [2]. This field of research does not only push forward the development of machines but also helps to better understand human orientation. Path integration principles are applied with the navigation of blind people, patients with vestibular defects or in theoretical work [2].

Outlook
In order for proper and smooth orientation, remembering coincides with learning. Insects have to learn the position of a goal and the route to it. Further routes can also be learned from other individuals. Probably the best known example to show this behaviour are social bees (Apis mellifera) [5].

Despite of the three different orientation mechanisms, insects seem to lack a cognitive map. In fact it is arguable, whether or not there is evidence against such a map when considering path integration [5]. However, one should keep in mind that a cognitive map in insects could exist based on yet unknown mechanisms [5]. Learning and the discussion of cognitive maps have not been covered here, but may be of interest for further reviewing.