Tactical Edge Reprogramming for Rapid Autonomy Adaptation

Aloise, Megan

Student Work

Tactical Edge Reprogramming for Rapid Autonomy Adaptation

公开 Deposited

Tactical Edge Reprogramming for Rapid Autonomy Adaptation (TERRAA) focuses on developing robotic agents that can work as a member of human teams in tactical settings without the involvement of a robotics engineering expert. A subset of the TERRAA project, Tactical Formulaic Language Interpretation and Prediction (TFLIP), aims to achieve this by using Tactical Language, a formulaic language utilized by many human Special Weapons and Tactics (SWAT) and Army Special Reaction teams. This language is particularly relevant to tactical room searches, where US Army officers will use this language to describe a room. This language emphasizes the geometry and features of the room that are relevant to method of entry. The TERRAA TFLIP project aims to take advantage of this tactical formulaic language already understood by humans trained in room clearing and introduce that language to a robotic operative. The robot’s goal is to utilize the provided verbal description of the room and show an understanding of the description by predicting unseen features of the room. This will be achieved by combining real-time sensor readings and prior knowledge of the room’s geometry. The robot entering the room will contain multiple sensors which it can use to observe the room. These sensor readings represent incomplete observations of the room. The particle filtering algorithm is a common method in robotics to estimate a true state based on incomplete observations (Del Moral, 1996). Implementing such a particle filter for our application is a complex process comprised of many elements. One important aspect is determining the score of a particle provided the sensor observations. This project assesses how different methods of scoring and weighting of known information affect the accuracy in particle filter convergence toward the correct true room at various stages of receiving information. Due to the incomplete initial state of the project, other aspects of the particle filtering algorithm, such as particle representation, particle generation, and particle elimination will also be addressed. Unfortunately, due to the non-public nature of official tactical language dictionaries and room search protocols, existing autonomous projects created for tactical room search context are sparse to non-existent. However, the particle filtering algorithm has been used in various civilian applications which shall serve as inspirations for this project. In investigating various room representation and generation techniques, we evaluated a coordinate approach, a geometric properties approach, and a door-centric geometric properties approach. The door-centric approach was found to be the most efficient representation and generation technique due to its simplicity and unique relevancy to tactical language room descriptors and room entry perspective. In investigating various particle elimination techniques, we evaluated a deterministic percentile approach, a probabilistic draw approach, and a probabilistic individual score assessment approach. The probabilistic individual score assessment approach was determined to be the favorable approach due to concerns of improper convergence and lack of accommodation for multimodal score distributions with other techniques. The scoring functions were investigated by creating edge-case representative test cases and collecting room estimate data from each of four iterations of the particle filter for each test case. These were collected while running the particle filter on each separate attribute’s optimization score contribution. By doing this, it was determined that three of the six attempted optimization techniques were effectively isolated, where two of the three unsuccessful optimization techniques showed irregular success. The three unsuccessful optimization techniques require further investigation. Once each attribute is determined to have a successful optimization technique, the method of combining each of these score contribution values will require investigation. In addition to attribute optimization, it was discovered from gathering these results that particle generation will need to be altered to accommodate for all potential room attribute values. These findings have identified a promising room representation and generation technique, identified a promising elimination of particles technique, and evaluated individual attribute optimization scoring functions to pave way for future developments of the project.

This report represents the work of one or more WPI undergraduate students submitted to the faculty as evidence of completion of a degree requirement. WPI routinely publishes these reports on its website without editorial or peer review.

Creator