My activities this week have been primarily focussed on continuing the insect navigation research to see if I can pin down some core principles, though some research into the available simulation tools has also been undertaken, to determine an appropriate environment to use.
The key issues arising from the research appear to be that navigation capabilities differ between species, hopping between landmark identification and scene-matching depending on the research, so there is not a one-size-fits-all solution.
However, there are some common features identified in the research:
- Optic flow is used to monitor speed, aiding in the determination of distance. In bees, this also appears to be used in their centring response.
- Path integration is used to recall distance and direction.
- Physical saccades (peering behaviours) are used to judge distance when not in motion, due to the insects fixed eyes and focal point.
- They undertake two forms of navigation: way-finding (point A to point B over non-local distances with recognition of several places) and local (goal recognition over local distances).
- Basic pattern recognition is used during local navigation and, for bees, orientation of geometric shapes is of less relevance.
It was also noted that ants cannot scale a scene snapshot, so need to get an exact match before changing direction, whilst bees appear to be capable of scaling the snapshot and calculating the angular difference to determine the heading required.
Implementations in robots appear to focus on the extraction of landmarks from the scene and building up an environmental map consisting of vectors used to locate them.
Attempts to used simulation software this week has proven harder than anticipated. Whilst Flame appears to be a powerful tool, information is sparse and would lead to a steep learning curve, for which there is no time available. Additionally, attempts to get MASON working proved fruitless, leading to the consideration of changing the direction of the project to work on the Darwin OP robot used by the Robocup team.
The plan for the coming week is to:
- Investigate the DarwinOP software to gain familiarity and start building a quick and simple prototype. The purpose of using this environment and existing Robocup software would allow for the model to be tested in software and, if successful, within a physical robot.
- To produce a simple prototype in any multi-agent simulator I like to work out the algorithm for calculating the speed of landmarks passing an agent and, therefore, the distance to them.
As with the previous week, my main concern is with the mathematical understanding required to achieve the above.