The Parameter Space Illumination (PSI) library uses the concept of quality diversity optimization to improve the performance of autonomous robots. In the context of autonomous navigation, the paths to follow are generated based on sets of Parameters-Value Sets (PVS). Normally these values are adjusted by an expert and only one PVS is used. The PSI library produces a certain number of different, well performing PVSs for a robot to choose from.

The PSI library finds PVSs by running simulated missions with the robot in environments that are increasingly difficult to navigate and automatically estimating the performances. The algorithm enters these PVSs and their correspondent performances into a so called Behavior Performance Map (BPM). The BPM aims to cover a certain area of a 2D space defined by 2 of the parameters to be optimized, which define the Low Dimensional parameter space (LD). The PSI search is done in a discrete space. Thus, the BPM is a grid map filled with PVSs - arranged according to the LD - linked to their performance. The BPM only includes the PVS with the highest estimated performance, for each possible combination of the LD. The algorithm yields in this manner a set of solutions spread along the LD, which are locally competing with other solutions - where the LD are equal.

The PSI library already exists and the goal of my thesis will be to write and perform a test mission to evaluate if and how much this approach improves the performance of an autonomously navigating rover compared to the case of having a single setting adjusted by an expert.
The test mission will be one long route through simulated terrain that the robot will attempt in two different ways. The first way is where the parameter space illumination library will be used. The long path of the mission will be divided by a number of way-points. And the robot will have access to the BPM produced beforehand by the PSI. It will traverse until the next way-point using the PVS from the map with the highest performance. Then, at each way-point the previous performance estimation for the used PVS will be recalculated taking into account the performance in the last traverse and updated in the bpm. Then the robot will traverse to the next way-point, selecting and using the best performing PVS from the updated BPM. The second way will be on the same path, but with only one set of parameters which will be hand tuned by an expert and used during the whole way.

Then I will compare the duration and encountered navigation failures with and without the BPM, to produce quantitative results. I will run the mission multiple times to gain statistical significance.
I will implement the functionality needed to run the test mission as an addition to the library, gaining a deep understanding of the current code to see which components can be reused for it and where changes need to be made.
I will use UML diagrams to plan these changes before starting with the implementation.
I will also choose from a selection of simulated landscapes and design the path for the mission, with the goal of making the test as effective in evaluating the performance of the library as possible. Then I will run the test, analyze its results, produce visualizations summarizing these and draw a conclusion about them.