Explore Projects

This project aims to significantly enhance two specific simulation scenarios within the JdeRobot Robotics Academy: Follow Line and Vacuum Cleaner. While these exercises serve as valuable educational tools for robotics learners, the current Gazebo environments and robot models lack visual realism, accurate physical behavior, and modular ROS2 integration. To address these limitations, the project will refactor the URDF model used in the Follow Line scenario using modular Xacro descriptions and realistic textures. The line track world will also be improved with high-quality road textures and better lighting for consistent perception results. In the Vacuum Cleaner scenario, the environment will be rebuilt using realistic room furniture and obstacles imported from Blender assets. Additionally, the Roomba robot model will be upgraded for improved physical dynamics, and its SLAM and navigation performance will be enhanced by tuning ROS2 Navigation2 parameters. Common improvements across both scenarios include updating the launch files to align with ROS2 Humble best practices, ensuring compatibility with the latest versions of Gazebo such as Fortress or Ignition, and producing clear documentation to support scenario creation and extension. The final deliverables will include the revamped simulation scenarios, reusable robot models, an organized and modular launch structure, and comprehensive user documentation. These improvements aim to boost the realism, engagement, and overall educational value of the Robotics Academy platform.

This project proposes the development of a foundational model for symbolic regression tailored to high energy physics (HEP). Symbolic regression can be used to rapidly provide solutions to scientific problems with high computational complexity or even intractable tasks, discovering symbolic expressions that describe underlying physical laws. While many approaches have been explored, this project focuses on building a foundation model which incorporates knowledge about the task of symbolic regression and common data representations for physical systems. The model will target accurate and interpretable prediction of squared amplitudes and cross sections—two core quantities in particle physics. Leveraging physics‑informed representations, symbolic machine learning, and simulation data, the project will explore how to integrate domain knowledge into learning algorithms to enhance scientific validity. Over 28 weeks, the project will progress through phases of theoretical investigation, foundation model design, simulation‑based data generation, model development, and integration. Deliverables will include a curated dataset, pre‑trained foundation model, symbolic models, and a final report, with broader goals of improving interpretability and reliability in physics‑informed AI.

The CHAOSS community develops metrics, practices, and software for measuring open source software community health and sustainability. Augur and Grimoirelab are two powerful open source community health analytics projects supporting the CHAOSS mission. In particular, each has implementation of select metrics based on an expanding set of implementation-agnostic metric definitions associated with various focus areas. The CHAOSS Metrics Model Working group defines metric models, which integrate multiple CHAOSS metrics in a way that makes it easy for end users such as community managers to get a more comprehensive measure of an open source community. Currently, Conversion Rate can be defined as a metric or metric model. It measures the rates at which new contributors to an open source project move through the "contributor funnel" toward sustained contribution and/or community maintainership. There is no implementation of the Conversion Rate metric or metric model in CHAOSS software yet. This project aims to build a novel implementation of Conversion Rate as a metric model that is based on a minimum number of atomic metrics and has maximum reuse of existing implemented metrics as well as common metric model calculation code supporting all types of conversion rates. A simple config file based UI for entering parameters and filters by the users is all it takes to run the metric model. New metrics will be implemented first in Jupyter Notebook as reference code and can later migrate into Grimoirelab backend (for sharing with future new metric models or reports) along with data collection and enrichment extensions. The Metric Model calculation code and methods, and a common visualization interface will be also implemented in Jupyter Notebook, followed by a deployment with OpenSearch Dashboards visualization.

<p>Icarus is an open source tool used for light curve synthesis. The aim of the project is to add more functionalities to Icarus.</p>

This project extends the Fisheries Integrated Modeling System (FIMS) to support user-defined sex structure in population models. Currently, FIMS assumes a fixed 50/50 sex ratio, limiting its ability to represent sex-specific population dynamics used in stock assessment. To address this, the project introduces a partition-based framework that can enable sex-specific derived quantities (e.g. spawning biomass, landings) and population dynamics (e.g. mortality, selectivity) while minimizing computational overhead. Rather than implementing sex structure as a standalone feature, this approach supports a more general and extensible design for introducing partitions into the model. Three levels of model complexity will be developed and evaluated to balance flexibility and performance. The ultimate goal of this project is to implement an explicit two-sex model, which enables the capacity to model life-history dynamics differently between different sex groups. Benchmarking experiments will assess runtime and identify efficient configurations. Upon completion, this feature will expand the applicability of FIMS to a broader range of fisheries and stocks across the United States.

<p>An interesting study is to construct a model of the media representations of the world, considering features from social discourse such as crime, race, and so on. In other words, we generate “interpretive frames” that introduce selected media biases and predispose the system to look at the world in a certain way. Therefore, any data that is presented to this system will be viewed through this lens where certain outcomes are anticipated, and the communicative effects will depend on the associated inferences.</p> <p>We propose a few models and studies:</p> <ul> <li>Model 1: Generating different summaries of a news story by opinion spectrum based framing</li> <li>Model 2: Generating parody news stories based on an image using neural storytelling</li> <li>Model 3: Simulating evolving window of acceptability</li> <li>Model 4: Simulating the spread of unverified information</li> <li>Study: News recommender systems causing filter bubble effect</li> </ul> <p>With the help of these, we may conclusively demonstrate real world phenomena that highlight issues of media bias, framing, echo chambers, and the wisdom of the crowd. Such an analysis may shed some light on the political climate and how crowd consensus comes to be.</p>

<p>TensorFlow Model Garden and TF Hub provide a lot of state-of-the-art models for TF developers to use easily in their research and product development, so, the key feature in these tools is to provide state-of-the-art models and this is what the project is about. So, the mission of the project I’m applying for is to join a team on the quest of training new SOTA models and publishing them. For this, I have devised this problem into 2 sections:</p> <ul> <li>Training and publishing SOTA models: This main part consists of working with the team on creating, training, documenting, and publishing the models that they have planned on adding. This consists of my contribution to the architecture and implementation of these models.</li> <li>Proposing models to add: I am making sure I keep my innovative side in this program by trying to suggest ideas of models that can be added to TF Hub. For instance, there are few models that treat underrepresented languages in NLP, like Arabic. My contribution to this can be crucial as I have experience dealing with these types of models and am aware of the difficulties someone can encounter when dealing with this. This can also be generalized in other fields, not only NLP</li> </ul>

<p>The Functional Mockup Interface (FMI) is a tool independent standard for the exchange of dynamic models and for co-simulation. The main objective of this project is to create an interface for model exchange and perform co-simulation.</p>

The objective of the project is to implement a new state-of-the-art 3D Object detection model (PV-RCNN++) in PyTorch and TensorFlow and train it on large-scale datasets such as Waymo to reproduce the original author’s results. The expected outcomes of the project are: • Implement PV-RCNN++ training and evaluation with PyTorch. • Train and reproduce results on the Waymo dataset. • Port the PyTorch implementation to TensorFlow and ensure results can be reproduced. • Proper documentation about model usage and config customizations with tutorials and examples. The implementation plan is as follows: • Using OpenPCDet implementation as a reference, implement PV-RCNN++ in PyTorch and benchmark on Waymo Open Dataset. • Using this implemented PyTorch model as a reference, implement each corresponding layer in Tensorflow and test it. Convert the PyTorch model to TensorFlow and verify the metrics on the Waymo dataset. • Once verified, train the model from scratch in TensorFlow and verify the results.

This project aims to extend the capabilities of the ONNX-to-mlpack converter by adding support for more advanced and diverse model architectures. The focus will be on enabling the translation of complex ONNX computational graphs, including LSTM-based models and Directed Acyclic Graph (DAG) structures. Additionally, the project will introduce support for NLP models and incorporate foundational NLP components such as Word2Vec-based tokenizers, encoders and decoders in mlpack.

This project aims to enhance the performance of the LLAMA2 model by fine-tuning it specifically for Jenkins data, thus enabling users to access Jenkins knowledge through a user-friendly interface. By leveraging data from Jenkins documentation and other relevant sources, the model will be refined to better understand and respond to domain-specific queries. The project involves conducting data preprocessing, fine-tuning the LLAMA model, designing a simple graphical user interface (GUI), and integrating the refined model into the system. Through testing and user feedback, the project aims to ensure the robustness and usability of the system. The deliverables include a system architecture design document, data preprocessing/refining documentation, a user-friendly GUI, a fine-tuned LLAMA2 model, testing reports with results, and a maintenance plan/documentation. The project will empower users to interact with Jenkins knowledge more effectively, thereby enhancing the accessibility and usability of the software.

Convolutional Neural Networks require large receptive fields in order to track long-range dependencies within an image, which in practice involves large kernels or long sequences of convolutional layers at the cost of losing efficiency. The problem can be mitigated by using Equivariant Transformers. Attention layers can directly replace the convolutions, so they will be able to attend to a larger receptive field than regular convolutions. In addition, the equivariant part of the model can replace the manual augmentation applied on every batch during training.

<p>Stochastic variational inference is a powerful tool for analyzing probabilistic models, especially for large scale problem. In this project, our goal is to implement a generic algorithm based on stochastic variational inference for a large family of probabilistic models, including latent Dirichlet allocation (LDA), hidden Markov models (HMM), and Bayesian mixture models (BMM) and etc. We describe the design and the implementation plan of our project and propose a timeline for our development.</p>

For Historical Manuscripts, limited availability of data a limiting factor in training OCR or Layout Detection models. In this project I propose the WGAN architecture to generate synthetic renaissance style Image Data which can be used to improve OCR model performance on historical Spanish texts.

<p>This project aims to create a web based code editor for Information models provided on Eclipse Vorto. The web editor will be used to provide similar functionality and user experience for editing Information Models as provided by the Vorto plugin on Eclipse. The Information models created will be stored in the central JCR repository but can only be edited by the users who created them. Through this project I aim to implement this enhancement : <a href="https://github.com/eclipse/vorto/issues/66" target="_blank">https://github.com/eclipse/vorto/issues/66</a></p>

<p>This project aims to build and implement a Data model for logs in a large kubernetes cluster to process, correlate and query to make the troubleshooting easier and can reduce the time in finding root causes.</p> <p>The objective of this proposal is to twofold.</p> <ol> <li>Build on the <a href="https://github.com/ViaQ/elasticsearch-templates" target="_blank">existing</a> work on Data model from Red Hat and others in the open source community, demonstrate the work and identify any gaps. </li> <li>To enhance the data model to based on the gaps.</li> </ol>

For the last 200 years research into human cognition and decision-making has revolved around rational choice theory, which assumes that actors are utility maximizers capable of searching and finding rational and optimal decisions. However human behavioral data doesn’t support the assumptions that we have infinite time nor infinite cognitive energy to search the full space of candidate choices to produce optimal decisions. Take for example a hiker, who has an infinite number of paths to choose from. It's irrational to assume that the hiker is capable of fully evaluating each of the infinite paths available to them to then choose the most rational/ optimal path. And yet that is the operating assumption for most modern cognitive models. To remedy this problem, I propose three axiomatic principles of energy-efficient decision making, along with an optimal-control based model and Deep Q learning model that capture those principles. I hope that a formalization of my theory of energy-efficient decision making along with the open-source python library we develop can be used by researchers in private and academic settings to develop more accurate computational models of human decision making, that can be used to model agent behavior in a number of fields ranging from market behavior to crowd dynamics, along with improving LLM infrastructures.

<p>There are many alternatives for L1-regularized generalized linear models in R, but none that utilizes the efficient SAGA algorithm despite its excellent convergence properties and track record of compelling empirical results. A successful implementation of the SAGA algorithm does, however, exist in <strong>scikit-learn</strong>: an actively maintained and well-documented Python module. The goal of this project is to port that implementation to R as a package targeted for submission to the Comprehensive R Archive Network (CRAN).The end result will be an easy-to-use and blazingly fast algorithm for L1-regularized models, which will come as a wanted addition to the toolkit of R users interested in big data modelling.</p>

The project's core objective is to develop an open-source foundational model for EEG data analysis, using deep learning techniques and extensive pre-training on a broad spectrum of EEG datasets. This foundational model will enable more effective processing, feature extraction, and interpretation of EEG signals, catering to both large-scale datasets and specific, smaller datasets.

<p>Ordinary differential equations are widely used for modelling biological networks and processes of kinetics. The idea of the project is to create a web application that allows a user to submit an SBML model and select methods of analysis and receive results. The different kind of analysis methods that will be available for use are Steady State Analysis, Parameter Fitting, Simulation.</p>

KubeEdge, as a native edge computing platform built on the Kubernetes ecosystem, offers capabilities such as reliable cloud-edge communication, edge autonomy, and IoT device integration. However, its ability to support intelligent model execution at the edge has yet to be systematically validated and practiced in real-world scenarios. This research aims to explore the feasibility and performance of deploying and running small language models on edge nodes using KubeEdge Expected Outcome: 1. Verification of KubeEdge's model deployment capability at the edge. Deployment and testing of model engines such as vLLM and llama.cpp on edge nodes, along with providing practical examples and detailed documentation for deploying small language models. 2. Exploration of integration schemes between KubeEdge and the OPEA platform. Connecting KubeEdge with OPEA’s model registry and workflow orchestrator to support automated model distribution and deployment from the cloud to edge nodes.

Meshery Models are declarative representations of infrastructure and applications. Within these models, Relationships define how different Components (e.g., Kubernetes resources, Cloud services) interact and depend on each other. These relationships are crucial for visualizing, understanding, and managing complex cloud native systems. This internship focuses on significantly expanding the breadth and depth of Meshery Relationships across a wide array of technologies supported by Meshery. As Meshery continues to integrate with more cloud-native technologies (Kubernetes, public clouds, and all CNCF projects), there's a growing need to accurately model the intricate relationships between their components - vital for providing users with comprehensive insights and control over their deployments. Expected Outcome: - A multitude of new relationships defined both intra and inter AWS services. - Policy Contribution: For advanced interns, there may be opportunities to contribute to the Rego policies that evaluate and enforce these relationships.

Meshery Models are declarative representations of infrastructure and applications. Within these models, Relationships define how different Components (e.g., Kubernetes resources, Cloud services) interact and depend on each other. These relationships are crucial for visualizing, understanding, and managing complex cloud native systems. This internship focuses on significantly expanding the breadth and depth of Meshery Relationships across a wide array of technologies supported by Meshery. As Meshery continues to integrate with more cloud-native technologies (Kubernetes, public clouds, and all CNCF projects), there's a growing need to accurately model the intricate relationships between their components - vital for providing users with comprehensive insights and control over their deployments. Expected Outcome: - A multitude of new relationships defined both intra and inter Azure services. - Policy Contribution: For advanced interns, there may be opportunities to contribute to the Rego policies that evaluate and enforce these relationships.