48440 research outputs found
Sort by
Global State Prediction for Reinforcement Learning in Collective Transport
No full textThis dissertation introduces Global State Prediction (GSP), a novel framework addressing the fundamental challenge of non-stationarity in multi-agent reinforcement learning for collective transport tasks. The most significant contribution of this work is a paradigm shift in how we conceptualize coordination in swarm systems—moving from information-sharing to prediction-based approaches. Our research demonstrates that enabling agents to anticipate collective outcomes can be more effective than traditional methods, even when using significantly less communication bandwidth. The GSP framework enables decentralized coordination through a prediction mechanism where agents anticipate future global states based on shared partial observations. Our extensive experimental evaluation demonstrates that GSP consistently outperforms both implicit communication baselines and, surprisingly, global knowledge approaches that utilize complete state information. Building on this foundation, we introduce Neighborhood-Limited GSP (GSP-N), which restricts information exchange to immediate neighbors while maintaining performance comparable to broadcast approaches, reducing communication complexity from O(n2) to O(1) as swarm size increases. We further extend this framework with two memory-enhanced variants that incorporate temporal reasoning: Recurrent GSP-N (R-GSP-N) leverages LSTM networks to process historical interaction patterns, while Attention-based GSP-N (A-GSP-N) employs transformer architecture to selectively focus on relevant information across both space and time. These approaches establish a clear performance hierarchy that challenges conventional assumptions about coordination requirements in multi-agent systems. Our work uniquely addresses the transport of non-uniform objects with unknown mass distributions, with GSP variants demonstrating remarkable resilience where baseline approaches experience significant performance degradation. We validate our approach through successful sim-to-real transfer experiments with physical robots, demonstrating emergent role specialization and robust coordination without requiring policy modifications. Notably, policies trained with GSP-N variants transfer effectively to larger swarms without retraining, demonstrating exceptional scalability. This research advances both theoretical understanding of multi-agent coordination and practical capabilities for robust collective transport in real-world applications
An Engineer's Tour of Iceland's Renewable Energy
No full textIceland is widely known for its stunning natural landscapes, but its significant contributions to science, technology, engineering, and math (STEM) remain underrepresented in tourism. This project aims to design a STEM-focused tour in Iceland that highlights the country's unique technological advancements, sustainable energy initiatives, and industrial innovations. By identifying potential tour locations, assessing tourist interest, and creating a comprehensive itinerary, this project aims to promote educational tourism while alleviating environmental strain on Iceland's current tourism locations, which are largely nature-based. The project's goal is to highlight Iceland's leadership in STEM and explore sustainable strategies to support the long-term viability of its tourism industry
Creating Trail Training Videos for the AMC
No full textThe goal of this project was to standardize and elevate sustainable trail maintenance practices. In collaboration with Appalachian Mountain Club's (AMC) trail program team, we produced training videos for use by AMC, volunteers, and the public. We utilized an iterative design process to refine the videos using our qualitative analysis of 15 instructional videos, feedback from 5 interviews and 28 survey respondents. We developed four videos: titled "Ski Trail Brushing", "Intro to Trail Blazing", "Water Drainage Maintenance", and "Trail Tools Safety". We recommend that the AMC leverage our findings on producing engaging educational videos, and the tools that proved most helpful, to continue to expand their video catalog and their outreach to a new generation of volunteers
A Smart Catheter System for Catheter Clot Interaction Monitorization of Aspiration Thrombectomy
No full textIn 2020, the global prevalence of stroke reached 89.13 million cases, with acute ischemic stroke (AIS) accounting for approximately 68.16 million cases. AIS occurs when a blood clot obstructs blood flow to the brain, often leading to severe neurological damage. Mechanical thrombectomy using aspiration catheters has become a standard intervention; however, current techniques lack feedback to assess catheter-clot interactions, increasing the risk of incomplete clot removal or vessel collapse. This project developed and validated a physiologically relevant benchtop model of the intracranial environment to investigate these interactions using a SMART catheter integrated with a novel fiber-optic pressure sensor. Clot surrogates, artery surrogates, and a benchtop model simulating the intracranial pressure environment were used to replicate three common thrombectomy scenarios: successful clot suction, clot jam, and arterial collapse. Pressure trends captured by the fiber-optic sensor integrated in a standard Penumbra aspiration catheter confirmed distinct interaction profiles across these scenarios, demonstrating the feasibility of our SMART catheter for clot engagement monitoring. Validation of this integration supports the potential of fiber-optic sensing to enhance the safety, precision, and efficacy of aspiration thrombectomy procedures for stroke treatment
A Unified Framework for Interference Mitigation and Cooperative Beamforming in Vehicle-to-Vehicle Networks
No full textThe advancement of CAVs has intensified the demand for resilient and scalable V2V communication in complex, dynamic environments, where challenges such as sensor uncertainty, positional inaccuracy, wireless fading, and co-channel interference persistently complicate safe and efficient cooperation. This dissertation presents a comprehensive framework that integrates realistic mobility modeling, robust state estimation, adaptive communication, and decentralized consensus mechanisms to effectively address these issues. A modular end-to-end simulation platform is developed, integrating spline-based road geometries, Markovian lane-switching, and detailed environmental effects such as weather-dependent visibility and variable road friction. The vehicle state estimation is achieved with the UKF, allowing compensation for GPS errors and enabling adaptive beamforming strategies including null-steering and interference suppression to maximize SINR in dense, interference-prone conditions. A trust-weighted consensus protocol for cooperative merging combines real-time PDR with historical trust, supporting decentralized leader election and robust decisions despite V2V uncertainties. The quantitative analysis shows that the proposed approach significantly enhances estimation accuracy, safety margins, and CA relative to fixed-leader or static-rule baselines. By integrating reliability-aware consensus mechanisms, cross-layer feedback, and adaptive trajectory planning, the framework establishes a scalable foundation for next-generation cooperative driving and ITS. The results and methods provide guidance for designing and deploying robust, reliability-aware protocols in real-world vehicular environments, supporting future advances in autonomous platooning, dynamic merging, and networked transportation infrastructures
CHIPNet: Coarse-to-Fine Hierarchical Inference for Precise Corner Detection on Chips using Neural Networks
No full textThis thesis presents CHIPNet, a three-stage hierarchical deep learning framework for precise corner detection in semiconductor chip images. The system achieves 99.06% patch detection accuracy and localizes 93.87% of corners within 15 pixels (median error: 5.71 pixels), significantly outperforming traditional computer vision methods and direct deep learning approaches. Developed in collaboration with Teradyne, the framework demonstrates strong generalization across different chip designs and is delivered for deployment in production semiconductor testing facilities
IQP 2517 Stock Market Simulation
No full textThis Interactive Qualifying Project (IQP) evaluates short-term stock trading strategies through a six-week simulation using the Investopedia Simulator. Five stocks across diverse sectors were analyzed with a swing trading approach based on moving averages and Relative Strength Index (RSI). The technical simulation returned +8.24%, while a fundamentals-based simulation, guided by earnings, valuation, and news, yielded +3.87%. Comparing the two highlights how technical and fundamental strategies can produce different outcomes, offering insights into market trends, trading discipline, and decision-making challenges
Lead in Drinking Water of MA Education Facilities
No full textLead is a dangerous contaminant that impairs many functions of the body when consumed and is a problem in drinking water. We worked with MassDEP to conduct outreach to determine if and how education facilities addressed lead in their drinking water. During outreach, we found that childcare facilities responded more than schools, and we got more responses from emails than calls. From the responses, we learned about common remediation actions taken by facilities, like posting signage and installing filters. Due to our time limit, there are still many facilities to contact. We recommend that MassDEP prioritize more outreach using the toolkit we made and collaborate with other agencies to pass a bill mandating testing and remediation, so there is no question about how lead is being addressed
Portable Affordable Wheelchair Enhancer E25 MQP
No full textThis project investigates the reliability and efficiency of the Portable Affordable Wheelchair Enhancer (PAWE) system. The PAWE is a portable and lightweight device that is attachable to the back of a standard manual wheelchair. The focus of this project is to innovate previous PAWE designs, and to evaluate the durability, reliability, and effectiveness of this new design. The design prioritizes the compatibility with foldable standard manual wheelchairs, minimal structural modification, and ease of use
Stock Market Simulation 2512
No full textThe goal of this project was to compare two trading strategies through simulation to determine which would yield higher profits, in addition to building a deeper understanding of the stock market and developing skills for managing a personal investment portfolio. The strategies tested were a Technical Trading approach and a modified approach that incorporated Social Media Sentiment analysis to adjust and enter positions. Both simulated portfolios began with $100,000 in cash. Results showed that Technical Trading achieved a 10.2% profit, outperforming the Media Sentiment strategy, which returned 2.6%. The overall gain from the S&P 500 was 4.84% over the duration of the 6-week simulations. The simulation showcased that Technical Trading was more profitable than the Social Media Sentiment trading strategy. This project provided participants with practical trading experience and insight on what can affect and impact the market, so that they can make more informed investment decisions in the future