RBE594 Robotics Capstone Final Presentations
7:00 pm to 9:00 pm
Join us Via Zoom: https://wpi.zoom.us/j/94271159293
An Integrated UAV Planning-Vision System for Woodland Search and Rescue
Zachary Helfer, Mariya Huffman, Naseem Shah, Kip Talman
Time is of the essence for thousands of Search and Rescue (SAR) missions occurring each year across America’s sixty-four national parks. Our project demonstrates an efficient and scalable UAV-based solution for forested environments that operates alongside human SAR teams to decrease search time. A simulation is employed to demonstrate the efficacy of our hiker detection system. The system plans and executes search paths over areas of interest. A vision pipeline runs state of the art tracking algorithms in real-time to quickly identify candidate victims. Our results underpin the utility of robots in these high stakes searches. Their adoption will continue to explode as similar systems gain the approval of real-world search and rescue teams.
Swarm Robotics for Urban Fire Search and Rescue: Sensor Fusion and Local & Global Planning
Kristoffer Hidalgo, Joseph Lombardi, Pierre Rafiq & Tabatha Viso
In 2021, in the USA alone, there were 3800 civilian deaths due to fire, including 78 firefighters. Fires are particularly deadly in urban and indoor settings due to the dynamically changing and high-risk environments, numerous unknowns, and the limitations faced by firefighters juggling multiple priorities. Especially critical in urban fire scenarios is the proximity of other buildings and vehicles, which heightens the urgency to promptly put out fires. By designing a solution that can autonomously and rapidly explore large unmapped buildings as well as locate and assist victims, we can save lives and enable firefighters to focus on fire control. We propose a swarm of semi-autonomous agents that explore and assist victims under the coordination of a central authority. The team’s contributions extend across 3 key areas:
Sensor Fusion — Presented is a sensor suite design and sensor fusion algorithm to combine various sensors for a single agent. A simulation environment was developed using Unity Development Platform to test and compare individual sensors and different fusions of sensors to detect obstacles, smoke, fire, and victims. The sensor suite algorithms were then optimized to utilize the different sensors depending on the state of the environment.
Local Planning — Presented is a derivation of the closest frontier algorithm to explore 100% of reachable indoor space. Several algorithms were contrasted for their computational efficiency and their adequateness to a swarm deployment. A custom simulation environment relying on a hybrid grid and topological map was developed for this work.
Global Planning — Presented is a multi-objective bio-inspired optimization algorithm that utilizes Pareto dominance to enhance the swarm's collective behavior. Our approach conducts separate, parallel searches for optimized solutions for various individual, conflicting objectives, followed by a selection of the single best solution across all criteria. Pareto front visualization was used across static and dynamic simulations for algorithm development and analysis.
Simplifying Robotic Home Assistance: Translating Human Intent into Robotic Action
Brigitte Broszus, Keith Chester, Bob DeMont & Clinton Williams
Millions of people requiring physical assistance reside outside institutional settings with full time assistance available. The idea of in-home robotic assistance has been elusive for decades due to the complexities of unstructured home environments and difficulty of human-robot communications. This project demonstrates an end-to-end concept for in-home robotic item retrieval for non-technical users. By employing Large Language Models (LLM)s to determine the user’s desired object, we achieve a simplified interface for users to task the robot and enable communication with a high-level planner to dynamically create and control the robot action. The high-level planner is also utilizing the gains of contextual awareness that these LLMs demonstrate to allow better understanding through implied heuristics to shape its planning. We demonstrate autonomous mapping of the simulated environment through a Simultaneous Localization and Mapping (SLAM) algorithm and utilize an off-the-shelf object recognition platform (YOLOV8) to identify objects in the simulated environment. From there, objects can be categorized and queried from a state management module. The task planner synthesizes the desired object, the obstacle map and the states of known objects to design tasks to seek the desired object, retrieve, and return to the user. All is demonstrated through a ROS2 simulation in Gazebo.