Current Fellows Projects
Novel Reinforcement Learning based Algorithms and Generative AI based Mixed Reality Toolkit for Online Convex and Non-convex 3D Packing Problem – (2024-present) |
Rahul Kumar Gop (IEOR), under the supervision of P. Balamurugan Abstract:- The online 3D packing problem represents a significant challenge in combinatorial optimization with direct applications in logistics and manufacturing. While traditional solutions have focused on efficiently arranging convex cuboidal items, a substantial research gap exists in handling objects with complex, non-convex geometries. This work presents 3DPAC, a novel Reinforcement Learning (RL) framework designed to address this challenge. Our approach utilizes a custom simulation environment built on the Gymnasium platform to model the online packing process. We employ a Maskable Proximal Policy Optimization (PPO) agent that learns to select optimal placement actions, comprising the choice of item, its orientation, and its 2D coordinates. A key contribution of our current work is the development of a robust environment capable of training agents that generalize across a diverse range of container dimensions by leveraging observation padding and environmental randomization. We have successfully implemented a comprehensive object rotation feature, which is fully integrated into the environment’s state, action space, and validity checks. Current efforts are focused on resolving numerical stability issues related to action masking, which is a prerequisite for stable, large-scale training. The primary objective moving forward is to extend this framework to accommodate non-convex geometries, a task that will require novel methods for object representation, collision detection, and stability analysis. Ultimately, this research aims to lay the groundwork for an innovative Generative AI-based Mixed Reality toolkit to facilitate human-in-the-loop optimization of complex packing scenarios. |
Design & Development of Digital Twin for Real-Time Decision Support – (2025-present) |
Sayantan Maiti (IEOR), under the supervision of Jayendran Venkateswaran Abstract:- MTP focuses on developing a high-fidelity, smart digital twin of the IIT Bombay campus, with the greater aim of extending the work to industrial systems. The digital twin is to be built using an agent-based discrete-event simulation model, in AnyLogic or any other simulation-based software and integrated with GIS layers for accurate spatial mapping. The initial phase involves modeling the campus dynamics — particularly traffic and mobility patterns — as a proof of concept. The later stage will focus on emulating realworld operations with high accuracy and with the help of real-time data via IoT sensors, cameras, etc. The long-term goal is to enable real-time monitoring, predictive analytics, and intelligent decision-making using machine learning (ML) or reinforcement learning (RL) techniques. What-if analysis is to be performed to evaluate alternative policies or configurations. Once validated on campus-scale systems, the framework will be extended to collaborate with industry partners in logistics and operations domains, enabling them to simulate and optimize real-world systems such as warehouses, manufacturing lines, or transport networks. |
Network design for trucking companies – (2025-present) |
Ayush Yadav (IEOR), under the supervision of Ashutosh Mahajan Abstract:- Logistics companies like Amazon, Delhivery, TCI etc have large trucking networks that move goods between various hubs. However, because of asymmetry in demand, many trucks may run without full load. This thesis explores optimization models for finding three-way and four-way cycles that ensure promised delivery dates are met and the costs of running trucks is lowered. Automatic tools for constructing such networks from in a green field and modifying current an existing network, both are considered. |
Real time high-throughput defect prediction of tiny defects in small objects over conveyor belt – (2025-present) |
Sabari Karthick S (IEOR), under the supervision of P. Balamurugan Abstract:- The project involves developing real-time high throughput AI techniques for identifying tiny defects over small objects (e.g. small tablets, small capsules, etc) in a fast moving conveyor belt. The project involves developing and deploying AI based computer vision-enabled techniques to identify the region of interest and to localize tiny defects that might be hard to detect by humans. The developed techniques will be directly deployed on a working system for detecting such defects in pharma industry. We plan to develop lightweight deep learning models which can be very efficient in training on diverse types of defects and are fast in inference too. The project also involves complete workflow design from image capture, image preprocessing, region of interest extraction, inference of tiny defects and replying to a hardware device to accept or ignore an object for next phases. |
Deep learning models for distribution shift-based style transfer with image prompts – (2025-present) |
Trivikram Umanath (IEOR), under the supervision of P. Balamurugan Abstract:- Natural conditions lead to several conditions that lead to distribution shifts in pixel amplitudes, frequencies and phase of computer vision based systems which are equipped for working under normal circumstances. Under such extreme distribution shifts, the capability of these vision systems become subpar. Hence it important to impart such knowledge into vision systems to make them understand the scenes and objects when encountering distribution shifts. For this purpose, we consider a few pre-trained vision language models and train them to quickly adapt to new distribution shifts which occur in nature. Such systems will be deployed in drone cameras to adapt to agriculture survey and agriculture monitoring in high altitude regions involving heavy cloud cover, mist, snow and fog. |
AI algorithms for detection of cyber attacks in supply chians – (2025-present) |
Himanshu Maurya (IEOR), under the supervision of Manjesh Hanawal Abstract:- Supply chains are the back bone of modern operations. With the advent of Industry IoT 4.0, the supply chain now heavily depends on data collected by various IoT devices to forecast the uncertainties. However, use of IoT increases the threat landscape of the supply chains as attackers can now target IoT and either disrupt the data collected from their or use them to launch malicious attacks. This work aims to develop AI based algorithms to detect supply chain attacks through IoT networks. Our framework will collect information related to tactics, techniques and procedures used by the attackers as per the MITRE ATT&CK framework (https://attack.mitre.org/) and build models to discriminate the malicious vs normal behaviour in the IoT networks. We will then use this models to develop AI algorithms that can detect malicious attacks based on the traffic patterns generated by the IoT devices. We thesis will focus, data logging, datasets generation, training and testing algorithms. The end goal is to come with a frameworks that observers the logs of the IoT networks and automatically detects if any malicious behaviour is observed. |
Training LLMs for natural language to query translations – (2025-present) |
Anubhav Binit (IEOR), under the supervision of Manjesh Hanawal Abstract:- In various fields of engineering, information is gathers and stored for knowledge building. This knowledge is useful to enhance the perform the system. However, the way this knowledge is stored could be different and to access it one may use different languages. For example, the knowledge may be stored in various databases like PGSQL and MangoDB, and to access their information different query languages may be required. This becomes a tedious task when the information generated in the network grows. Large Language Models have emerged as a potential tools to get the required information from various domain using natural language. They can translate the natural language into an appropriate quary language automatically and retrieved the required information. However, to achieve this, they need to trained to translate the natural language to query languages. In a way, the only programming language one needs to know is natural language! In this thesis we will train a language model to translate natural language queries to SQL queries. This will be help in supply chain operations, where lots of information gathered in stored in various databases, and to access it one need to write SQL queries. We will generate datasets to train a language model to translate natural language |
Digitalization of Supply Chains and Operations – (2024-present) |
Saimadhav S (IEOR), under the supervision of Jayendran Venkateswaran |
Smart Indoor Logistics – Enhancing Efficiency, Safety, and Well-Being – (2025-present) |
Siddhesh Madkaikar (IEOR), under the supervision of Saurabh Jain |
Past Fellows Projects
Pricing Contracts in Air Cargo Revenue Management – (2024-2025) |
Pulkit Khandelwal (IEOR), under the supervision of Ashutosh Mahajan Abstract:- This report investigates pricing and capacity allocation strategies in air cargo logistics, focusing on interactions between carriers and freight forwarders under uncertainty of demand. We analyze three models: centralized decision making, decentralized coordination without buy-back, and decentralized coordination with a buy-back mechanism. Closed-form expressions for expected revenues are derived and optimal capacity allocation strategies are explored using numerical simulations. The results show that centralized coordination produces the highest revenue, while the introduction of a buyback contract significantly reduces inefficiencies in decentralized settings, with a revenue loss below 1%. These findings offer actionable insights for contract design in logistics networks with asymmetric information. |
Replenishment Scheduling for Combat Logistic Forces – (2024-2025) |
Shubham Joshi (IEOR), under the supervision of Manjesh Hanawal Abstract:- The Fleet Management services conducts missions using combat ships, aircraft and submarines that operate together in mission-driven groups. These units require regular fuel replenishment, but returning to port frequently is not practical. To address this need, shuttle ships are deployed to deliver fuel at sea. This project presents a Mixed Integer Programming (MIP) model designed to generate efficient refueling schedules for these operations. The primary objectives of the model are to minimize operational costs, maintain inventory levels above safety thresholds and ensure uninterrupted mission readiness. To improve the model’s flexibility and scalability, two operational scenarios were introduced. In the first scenario, basic calculations are used to check whether combat ships can reach their assigned group using only their internal fuel. If they are found incapable, the optimization model is then applied to determine whether a feasible plan can be made using one or two shuttle ships. The second scenario assumes that combat groups remain stationary and the model is used to generate a full refueling schedule over the entire planning horizon. A user-friendly software interface has been developed to allow Fleet Management services planners to input mission data, configure ports and assets and visualize results through interactive maps and animations. The system successfully runs on the data and generates feasible, cost-effective and interpretable schedules, demonstrating its strength as a practical and reliable tool for mission planning and Fleet Management services logistics. |
Digital Twin Simulation – (2024-2025) |
Arun Singh (IEOR), under the supervision of Jayendran Venkateswaran Abstract:- The simulation and control of distributed systems in real-time environments are increasingly essential in domains such as transportation, automation, and logistics. This thesis, titled Distributed Simulation Modeling and Analysis for Complex Systems presents the implementation of federated communication and synchronization between simulation models using the High-Level Architecture (HLA) standard. A digital twin serves as a virtual representation of a physical system, enabling simulation, monitoring, and optimization of real-world operations.To explore the integration of digital twins in distributed environments, two distinct simulation setups were developed. The first setup connects a simulation software (AnyLogic) with an external Java-based application via PITCH pRTI, enabling real-time interaction and dynamic parameter control through HLA-compliant messaging. The second setup models end-to-end supply chain behavior using two AnyLogic simulations running as separate federates over a local area network. These federates represent different layers of the supply chain and coordinate through time-synchronized interaction over RTI. In both configurations, the HLA standard facilitates modularity, interoperability, and synchronized time advancement. The first setup demonstrates external control of simulation parameters and event-triggered communication, while the second setup models inter-tier logistics coordination between retailers, distributors, and manufacturers. Each federate executes autonomously yet progresses in logical time steps, ensuring accurate message ordering and causality preservation. A practical application is demonstrated through a supply chain case study where manufacturers, distributors, and retailers are each modeled as autonomous simulation agents. This approach supports real-time monitoring, predictive analytics, and distributed decision-making. Moreover, the federated architecture ensures data confidentiality and stakeholder independence while enabling collaborative simulation. The results validate the feasibility of HLA-based distributed simulation as a scalable framework for building digital twin systems in supply chain environments. |
Deep Learning for Real Time Visual Inspection of Pharma Pills on Conveyor Belt using Computer Vision – (2024-2025) |
Mayur Dhanawade (IEOR), under the supervision of P. Balamurugan Abstract:- Defect detection in pharmaceutical pills is critical to ensure product quality and patient safety, as flaws in color, shape, surface texture, or size may signal manufacturing issues. Although existing defect detection systems are available, they are often expensive customizable, and efficient computer vision-based tool for detecting and localizing defects in pills using deep learning techniques. The proposed solution processes images of pills to distinguish defective units from non-defective ones, with a strong emphasis on minimizing inference time. The approach is divided nto two parts: defect detection in the side view and in the top/bottom view of pills. For the side view, the initial segmentation strategy involves a two-stage pipeline using the Segment Anything Model (SAM) followed by a custom-trained U-Net model. To reduce computational overhead, this is later replaced by a single-stage segmentation approach using a YOLOv8-seg model fine-tuned on a custom-labeled dataset. This modification leads to a significant reduction in processing time—by approximately 65%. After segmenting the region of interest, defect detection is carried out using a series of image processing operations, with Gabor filters playing a central role. The system also includes adjustable parameters, allowing users to control the sensitivity or scale of defect detection as per their inspection requirements. The result is a fast, flexible, and affordable defect detection pipeline suitable for real-time pill inspection in pharmaceutical settings. For defect detection in the top/bottom view of pills, traditional image processing methods are inadequate due to the presence of engravings, which makes it difficult to isolate actual defects. To overcome this, supervised learning models are explored, but they require a labeled dataset, which is not readily available to us. A semi-automatic annotation tool is therefore developed. This tool first segments and removes the engraving and edge regions using a custom U-Net model trained on a dataset generated via rule-based techniques. Defects in the remaining pill surface are then detected using a Gabor filter-based method, similar to the side view approach. The annotated dataset generated through this pipeline is then used to train a YOLO model, which effectively detects defects of various sizes and shows promising generalization. Additionally, a contextual comparison-based method is introduced to enhance robustness across different pill types by comparing each test image with a defect-free reference of the same pill type. We have developed a fully functional tool with a user interface, where the user can upload pill images from any view (side or top/bottom). If defects are present, the tool detects and localizes them using bounding boxes; if not, it indicates that the pill is defect-free. The user also has control over customizable parameters to adjust the scale of defect detection. All functionalities are seamlessly integrated into the interface. |
Developing Optimal Warehouse Policies and Planned Interventions for Enhanced Sustainability using PetriNets – (2024-2025) |
Deepak Naik (IEOR), under the supervision of Priyank Sinha |
Opportunity of Small Batch Movement of Freight on Indian Rail Networks – (2024-2025) |
Rohit Kumar (IEOR), under the supervision of Narayan Rangaraj |