214
E. Kagioulis et al.
ilar way. That is, if the starting point is unknown, or the best match is deemed
poor, the algorithm needs to run a full memory scan in order to set the window.
However, this uncertainty of “where in the world am I?” along with “which of
my memories are useful now?” is the fundamental problem of any algorithm of
this nature, as by restricting matches one is, amongst other things, constraining
the information available to the algorithm. Thus, while a good starting position
and initial match could result in a good run for the agent, the opposite can
be true. Another limitation is in nest search, where the agent must precisely
home from all locations, rather than general route following. To ensure they can
find the nest, ants perform learning walks immediately after leaving it, in which
snapshots facing the nest from multiple locations are stored. In this case the
sequence information will not be important/useful as these memories should all
be used to find the nest. At that point a mechanism is required where the agent
switches off sequence information and compares across all memories.
In the work presented here, we have used a simple way of using sequence
information, aimed at showing the utility of incorporating temporal information
in a route navigation algorithm. Other methods of using sequence information
may ameliorate the aforementioned issues. We could for instance, update the
window such that the best previous match is at the middle or implement a
dynamic window length which would grow if matches were deemed poor [11].
However, it is perhaps more profitable to consider complex matching processes
and ways of incorporating temporal information. For instance SeqSLAM [17]
compares multiple memories within a window to multiple current views and can
thus be robust to some bad matches. Another option would be to use artificial
neural networks that encode temporal information [6,14] and use these to learn
a combined representation of sequence, or indeed position, and route snapshots
as in [22]. It is unclear how one would implement a neural network to encode
both image snapshots as well as sequence information, but incorporating PI
information might be possible, and is the subject of future work.
Our priorities in developing these models is to push the algorithms for use in
closed-loop, real world, natural environments as experienced by ants. Assessing
how temporal information is useful in such environments, and comparing it to
what we know about how temporal information is used by insects, could allow us
to not only improve robotic navigation but also illuminate insect neurobiology.
References
1. Ardin, P., Mangan, M., Wystrach, A., Webb, B.: How variation in head pitch could
affect image matching algorithms for ant navigation. J. Comp. Physiol. A. 201(6),
585–597 (2015). https://doi.org/10.1007/s00359-015-1005-8
2. Ardin, P., Peng, F., Mangan, M., Lagogiannis, K., Webb, B.: Using an insect
mushroom body circuit to encode route memory in complex natural environments.
PLoS Comput. Biol. 12(2), e1004683 (2016)
3. Baddeley, B., Graham, P., Husbands, P., Philippides, A.: A model of ant route
navigation driven by scene familiarity. PLoS Comput. Biol. 8(1), e1002336 (2012)
4. Bradski, G.: The OpenCV library. Dr. Dobb’s J. Softw. Tools (2000)
E. Kagioulis et al.
ilar way. That is, if the starting point is unknown, or the best match is deemed
poor, the algorithm needs to run a full memory scan in order to set the window.
However, this uncertainty of “where in the world am I?” along with “which of
my memories are useful now?” is the fundamental problem of any algorithm of
this nature, as by restricting matches one is, amongst other things, constraining
the information available to the algorithm. Thus, while a good starting position
and initial match could result in a good run for the agent, the opposite can
be true. Another limitation is in nest search, where the agent must precisely
home from all locations, rather than general route following. To ensure they can
find the nest, ants perform learning walks immediately after leaving it, in which
snapshots facing the nest from multiple locations are stored. In this case the
sequence information will not be important/useful as these memories should all
be used to find the nest. At that point a mechanism is required where the agent
switches off sequence information and compares across all memories.
In the work presented here, we have used a simple way of using sequence
information, aimed at showing the utility of incorporating temporal information
in a route navigation algorithm. Other methods of using sequence information
may ameliorate the aforementioned issues. We could for instance, update the
window such that the best previous match is at the middle or implement a
dynamic window length which would grow if matches were deemed poor [11].
However, it is perhaps more profitable to consider complex matching processes
and ways of incorporating temporal information. For instance SeqSLAM [17]
compares multiple memories within a window to multiple current views and can
thus be robust to some bad matches. Another option would be to use artificial
neural networks that encode temporal information [6,14] and use these to learn
a combined representation of sequence, or indeed position, and route snapshots
as in [22]. It is unclear how one would implement a neural network to encode
both image snapshots as well as sequence information, but incorporating PI
information might be possible, and is the subject of future work.
Our priorities in developing these models is to push the algorithms for use in
closed-loop, real world, natural environments as experienced by ants. Assessing
how temporal information is useful in such environments, and comparing it to
what we know about how temporal information is used by insects, could allow us
to not only improve robotic navigation but also illuminate insect neurobiology.
References
1. Ardin, P., Mangan, M., Wystrach, A., Webb, B.: How variation in head pitch could
affect image matching algorithms for ant navigation. J. Comp. Physiol. A. 201(6),
585–597 (2015). https://doi.org/10.1007/s00359-015-1005-8
2. Ardin, P., Peng, F., Mangan, M., Lagogiannis, K., Webb, B.: Using an insect
mushroom body circuit to encode route memory in complex natural environments.
PLoS Comput. Biol. 12(2), e1004683 (2016)
3. Baddeley, B., Graham, P., Husbands, P., Philippides, A.: A model of ant route
navigation driven by scene familiarity. PLoS Comput. Biol. 8(1), e1002336 (2012)
4. Bradski, G.: The OpenCV library. Dr. Dobb’s J. Softw. Tools (2000)