Seminar M.Sc Students
BOSA: Boundary Outward Synthetic Augmentation for Post-Hoc Out-of-Distribution Detection
Eran Tascesme, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Prof. Neta Rabin
Abstract: Modern neural networks often perform well on in-distribution data but may fail when encountering inputs from unfamiliar distributions. This research addresses the problem of out-of-distribution (OOD) detection using only the frozen feature space of a pretrained classifier. We present Boundary Outward Synthetic Augmentation (BOSA), a post-hoc and training-free framework that generates pseudo-OOD samples directly from in-distribution training features. BOSA identifies class-boundary samples using a multi-scale neighborhood criterion and moves them outward beyond the observed in-distribution support. Based on these synthetic samples, we introduce two lightweight detectors: BOSA-KNNRatio, which uses relative distances to ID and pseudo-OOD samples, and BOSA-MarginalDensity, which compares feature-wise density patterns. The method requires no auxiliary OOD data, no generative model, and no retraining. Experiments on lead benchmarks show that BOSA is especially effective for challenging near-OOD detection, providing a practical and competitive approach for reliable post-hoc OOD detection.
Bio: Eran Tascesme is a M.Sc. student in the School of Industrial & Intelligent Systems Engineering at Tel Aviv University, supervised by Prof. Neta Rabin. He was selected for the Engineering Faculty Honors Program, where he completed a research project with Prof. Neta Rabin two years ago. His research focuses on reliable machine learning in open-world settings, particularly out-of-distribution detection.
Optimizing Recharging Depot Location in a Robotic Delivery Service Extended by Public Transportation
Yanir Zadickario, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Dr. Mor Kaspi
Abstract: The rapid expansion of the global on-demand economy has significantly accelerated the deployment of Autonomous Mobile Robots (AMRs) for last-mile delivery, yet their practical efficiency remains severely curtailed by limited battery capacities and low travel velocities. We consider a service design that integrates robotic fleets with existing public transit systems, allowing robots to travel onboard transit vehicles to conserve energy and significantly extend their operational reach across dense urban environments. Within the context of this service, we introduce the AMR Depot Location Problem, focusing on the strategic location of infrastructure to maximize the benefits of such integration. To address this AMR depot location problem, we developed a two-stage solution approach that integrates a high-resolution simulation model with an approximated Mixed-Integer Linear Programming (MILP) optimization framework. The methodology utilizes a simulation model to capture operational complexities - including real-world transit schedules, battery characteristics, and stochastic demand, to generate a history of evaluated designs. A Calibration Model then performs parameter fitting on this historical data to determine an effective maximum coverage radius. This calibrated parameter is fed into an approximated MILP Location Model, which identifies optimal depot sites by minimizing a combined objective of fixed infrastructure costs, operational routing, and penalties for unserved requests. We validated this approach using a Tel Aviv metropolitan case study involving 1,250 potential locations. To benchmark our approach, we develop a Hierarchical Large Neighborhood Search (HLNS) algorithm that uses the simulation model for solution evaluation. The HLNS demonstrates that substantial improvements can be achieved by first focusing the search on intra-zonal decisions and subsequently refining solutions through localized inter-zonal location adjustments. At the same time, the approach highlights the challenges of applying state-of-the-art metaheuristics in settings where solution evaluation is computationally expensive. In contrast, the proposed approximation framework significantly outperforms the already strong solutions identified by the HLNS. This improvement stems from using the computationally intensive simulation model to evaluate only the most promising solutions selected by the approximated MILP location model. Ultimately, this study provides a scalable decision-support tool for designing transit-enhanced robotic delivery networks that minimize both capital infrastructure and daily operational expenses by leveraging existing urban transportation backbones.
Bio: Yanir Zadickario is a M.Sc. student. program in Industrial Engineering at Tel Aviv University, specializing in Analytics & Optimization. His research focuses on Optimizing Recharging Depot Location in a Robotic Delivery Service Extended by Public Transportation.
Truck and Robot Last-Mile Delivery under Uncertain Handover Times
Gil Sorani, M.Sc. student in the School of Industrial & Intelligent Systems Engineering
Advisors: Dr. Mor Kaspi
Abstract: The rapid growth of e-commerce has led to a substantial increase in parcel volumes, intensifying traffic congestion and air pollution in urban areas and increasing the need for scalable last-mile delivery solutions. A promising approach combines delivery trucks with Autonomous Mobile Robots (AMRs): trucks transport parcels to centralized locations, from which robots independently complete the final deliveries, leveraging the speed of trucks while reducing local congestion. A key operational challenge of this model, however, is that customers must be present to unlock the robots and retrieve their parcels, introducing uncertainty into delivery success. While much of the existing literature assumes negligible or fixed handover times, empirical evidence indicates a non-trivial failure rate due to customer absence, resulting in increased costs and operational inefficiencies.
To address this challenge, we develop an operational optimization model that jointly determines truck and robot routing decisions as well as customer-specific waiting times. The model aims to improve delivery success rates by reducing "double-handling" trips, defined as cases in which customers must be served again following a failed delivery, while balancing waiting times against their cascading effects on subsequent operations.
The proposed research follows four phases. First, the problem is formulated as a mixed-integer linear program (MILP) to solve small-scale instances, complemented by decomposition approaches for larger settings. This approach separates the problem into four hierarchical sub-problems: truck routing, assignment, robot routing, and scheduling. Each sub-problem is modeled mathematically and solved using custom-developed exact algorithms or heuristics. Second, an event-based simulation framework is developed to evaluate system performance and waiting strategies under dynamic, real-time conditions. Third, a rolling-horizon framework integrates the static stochastic model with the simulation to assess the robustness of planning decisions in a dynamic environment. Finally, the study investigates practical and easily deployable waiting policies.
Bio: Gil Sorani is a M.Sc. student . program in Industrial Engineering at Tel Aviv University, specializing in Data Science and Artificial Intelligence. His research focuses on addressing the 'Last-Mile Problem' in parcel delivery using Autonomous Mobile Robots (AMRs), while considering stochastic handover times.
