M´ etodos de Clasificaci´ on
5.4 Algoritmos del ´ Area de IR
This research was successful in meeting the overarching objective of verifying an autonomous UAS swarming algorithm through DT&E. A testing framework utilizing modeling, simulation, and flight test provided adequate data to verify the algorithm against the chosen metrics and identify improvements for future research. Simulta- neous validation of this framework aimed to address the broader goals and challenges outlined in Chapter I and to provide a reference for future DT&E of autonomous systems. The resulting testing continuum was specific to this algorithm and paired the validated framework with the empirical data from the test program.
Early verification of the existing algorithm used simulation and included various modifications to the rule set and code architecture. This process familiarized the researcher with the autonomous system under investigation and provided insight into the underlying control laws and expected swarm behaviors. Qualitative data gathered during this stage was later used to develop the test plan, incorporating previously used parameters into missions that covered the desired performance envelope. This test
plan was not all inclusive and tended to be conservative in design since this was the first live, open-air flight test of this autonomous swarming algorithm.
The data from flight tests provided a baseline of swarm performance and indicated that the algorithm was able to accurately control individual vehicles and promote swarming behavior in a variety of conditions. It also highlighted the need to test and further refine the algorithm to improve performance and address various anomalies that occurred during testing. This included the need for relative velocity damping of the commanded responses to avoid the oscillations and additional data outputs to the LCM logs.
Due to the limited number of flights and resulting data from flight tests, a definitive set of ‘ideal’ parameters was not identified. However, a suggested parameter list for stable swarming based on observations throughout this research, and the qualitative and quantitative results is provided as follows: rmin ≥ 5 m, vmax = 3 m/s, fmin ≥ 10,
b = 0.25, u ≥ 5 Hz, 0 < a ≥ −1. Further experimentation is highly recommended to explore the full state space of this autonomous system and improve the validated testing framework.
What improvements, if any, are required to improve the swarming ar- chitecture for testing?.
This question was incorporated to address the “design for test” challenge in Ahner and Parson (2016). A significant portion of Chapter III was spent answering this ques- tion through the use of simulation and formulation to verify the original algorithm. To summarize, the first significant change was the incorporation of UgCS in the SITL testing environment for easy visualization of swarm behavior. Second, architectural changes to the algorithm were made to group rule calculations at a lower level and aggregate the results in priority order into a single velocity vector at the macro level.
This change also enables a more straightforward implementation of additional rules in the future. A significant change to the rule equations was implemented to ensure proper rule adherence for the swarm. Lastly, the addition of the argparse function was incorporated into the code to consolidate versions, enable more straightforward parameter variation during testing and allow for live streaming of algorithm outputs.
What role does quantitative and qualitative assessment have in the verification of autonomous systems?.
Throughout this research, both quantitative and qualitative assessment played critical roles in verifying algorithm behavior. The use of visualization in SITL testing allowed for qualitative analysis during iterative testing of the algorithm. This pro- vided valuable insight into swarm ‘goodness’ and behavior in a variety of conditions used in designing the flight test plan. Additionally, this analysis technique was used during flight testing to gauge the performance of the swarm and the success of the mission since quantitative data was not immediately available. Many of the lessons learned resulted from a qualitative reflection on the testing process as a whole.
Quantitative analysis was fundamental in verifying individual vehicle and over- all swarm performance during flight testing. The ability to collect, analyze and plot data provided a basis for conclusive results that the algorithm performed as intended. Without quantitative analysis, it is impossible to state this same conclusion defen- sively. However, this analysis also provided insight into regions of poor algorithm performance and led to many suggested improvements for future revisions of this algorithm. Multiple discussions paired qualitative and quantitative results through- out Chapters III & IV to provide a comprehensive analysis and verification of the autonomous swarm.
What measure(s) or metric(s) can be used to evaluate swarm behavior?. Chapter III highlighted the test objectives, metrics, and required data chosen to evaluate swarm behavior. The chosen metrics included: swarm “goodness,” or average deviation from the ideal vehicle-vehicle distance (m), percent violation of Reynolds+ rule set, the communication time delay of LCM messages, and packet loss of LCM messages. Results using flight test data in Chapter IV provide numeric values for these metrics and also discuss additional observations and conclusions related to other measures of swarm behavior. The additional measures included the half-median distance, commanded and actual vehicle velocities, and other combinations of data that were useful in analyzing swarm behavior.
How well do these measure(s) or metric(s) quantify swarm behavior and algorithm performance?.
The chosen metrics identified in Chapter III were sufficient in quantifying swarm behavior and algorithm performance in some instances. However, visualization of the entire ‘picture’ of swarm performance was not possible with these metrics alone. Other metrics summarized in Table 12 were incorporated into the analysis to supplement the existing four and provide clarity for individual vehicle and swarm behavior.