Current alignment techniques, such as RLHF (Reinforcement Learning from Human Feedback), typically rely on a centralized set of "safety guidelines" that reflect the cultural norms of the model's developers. This often leads to cultural misalignment, where the LLM may inadvertently erase local nuances, apply inappropriate social taboos, or fail to recognize regional linguistic etiquette.
This thesis aims to move beyond "cultural awareness" toward Cultural Verification. The student will design a systematic framework to evaluate and verify whether an LLM’s outputs adhere to the specific socio-cultural norms, value systems, and historical contexts of a target demographic.

Key Responsibilities & Tasks:

The research will be structured around four primary pillars, beginning with the development of a comprehensive Taxonomy of Cultural Correctness to define the specific dimensions—such as social hierarchy, religious sensitivities, and gender roles—necessary for nuanced LLM evaluation. Building upon this foundation, the study will propose a Verification Methodology centered on a "Cultural Guardrail" architecture, utilizing both programmatic checks and human-in-the-loop oversight to validate model responses. To test the efficacy of this framework, a specialized Benchmark will be curated, featuring a localized "red-teaming" suite designed to expose cultural hallucinations and latent insensitivities. Finally, a Comparative Analysis will be conducted to measure the performance of leading models, including GPT-4, Llama 3, and Claude, against these proposed designs across two or more distinct cultural contexts.

In this project, you will contribute to the cutting edge field of robotic tactile perception. Working with a humanoid robotic hand equipped with highly sensitive vibration microphones, your goal is to give the robot the ability to "feel" and identify what it is touching. As the robotic hand manipulates a diverse set of everyday objects, the embedded sensors will capture complex acoustic and vibration signals. You will use this raw sensory data to develop an intelligent system capable of accurately recognizing both the specific object and its surface texture.

Key Responsibilities & Tasks:

- Experimental Design and Data Collection: Design and execute robust data collection protocols.
- Signal Processing: Preprocess the raw audio and vibration signals. This involves filtering noise, segmenting the data, and extracting meaningful temporal and frequency domain features (e.g., spectrograms) that characterize different textures and materials.

- Machine Learning Pipeline Development: Design, implement, and train artificial neural networks. The pipeline will take the processed acoustic signals as input and output the correct object and texture classifications.

- Benchmarking and Evaluation: Rigorously test your models. You will benchmark different neural network architectures and classical machine learning algorithms against one another, analyzing key performance metrics like accuracy, robustness, and computational efficiency.

In this thesis/project, you will create a multi-layer traversability map for an autonomous off-road mobile robot. You will use an RGB-D camera to build a robot-centric map with derived geometric features, including slope, step height, holes, roughness, etc. The map should be generic so that it can be used with multiple robots; also, the robot size and capabilities will be parametrized. You will use datasets for outdoor environments, including roadside, grasslands, and forest edges.

The tasks include:

- Generate 2.5D elevation maps.
- Derive Traversability Features (Geometry Layer), i.e., slope, step height, holes, roughness, etc.
- Create a Semantic Layer (from RGB) to distinguish between grass, dirt, asphalt, rock, vegetation, obstacle, etc.
- Define a robot-aware traversability scoring, considering the volume of the robot and its capabilities, to create a cost map for path planning.

Traditional teleoperation often fails when the operator moves out of a fixed sensor’s "sweet spot." This thesis aims to improve this by estimating the operator's arm pose relative to their own torso, enabling a more flexible and intuitive workspace.

Key Objectives:
- Research State-of-the-Art human pose estimation models for real-time 3D skeletal tracking.
- Develop a robust pipeline to extract 3D landmarks from RGB-D or monocular feeds.
- Modify the existing ROS control architecture to transform arm coordinates into a body-relative coordinate system

This thesis aims to improve teleoperation precision by automatically adjusting the hand-to-robot mapping based on the operator's unique anatomy. By using markerless tracking, we can estimate finger lengths in real-time to create a calibrated, intuitive control loop.

Key Objectives:
- Keypoint Tracking: Implement state-of-the-art markerless 3D hand tracking.
- Biometric Estimation: Calculate individual finger lengths from extracted keypoints.
- Control Integration: Automate personalized adjustments within a ROS-based teleoperation workflow.

In this thesis, you will derive road-shoulder traversability from stereo/LiDAR, capturing slope, curb, ditch, and surface roughness to enable safe planning near road edges, and validate predictions against real field logs.

The tasks include:
* Build a pipeline to fuse stereo depth and/or LiDAR into terrain primitives: slope, curb/edge step, ditch depth, roughness (e.g., surface variance, ESF/planarity).
* Convert primitives to a traversability costmap/risk layer with clear thresholds and uncertainty.
* Evaluate on off-road/roadside datasets (e.g., RELLIS-3D, RUGD) and replay field logs for end-to-end validation (near-miss/stop rates, path feasibility).
* Report metrics: ROC/PR for traversable vs. non-traversable, curb F1/AP, slope/roughness error vs. ground truth or survey labels, and planning success on replays.

In this thesis, you will build a multi-modal 3D voxel mapping stack (elevation + occupancy/ESDF with semantics) for lane-less roadside terrain, producing uncertainty-aware traversability that outperforms 2D costmaps.

The tasks include:
* Integrate stereo/LiDAR (+ optional IMU/GNSS) into a real-time voxel map (TSDF/ESDF); maintain map updates under motion.
* Fuse semantic cues (grass/curb/obstacle masks) into elevation/occupancy; model uncertainty for risk-aware planning.
* Validate planner-ready layers (ESDF, traversability) and benchmark vs. 2D baselines on roadside datasets.
* Run ablations on depth sparsity/noise, voxel resolution, and update rates.

In this thesis, you will design a continual learning pipeline to handle distribution shift (new invasive grass types, seasonal/region changes) without catastrophic forgetting. Training occurs offboard (in the cloud/lab) with periodic rollout to robots; the system must propose new classes and preserve prior performance across the entire fleet.

The tasks include:

* Build a class-incremental pipeline (rehearsal/coresets, regularization, knowledge distillation) with reproducible splits (old vs. new classes).
* Add novelty/OOD detection to flag potential new species; integrate weak labels and active learning loops.
* Define metrics and dashboards for pre/post deployment (mAP on old classes, mAP on new classes, average forgetting, calibration).

In this thesis, you will develop a robust vision method for recognizing invasive grass patches under challenging lighting conditions, weather conditions, seasonal changes, and motion blur—mitigating domain shift via domain generalization and test-time adaptation.
The tasks include:

* Curate a multi-condition dataset (day/night, sun/cloud, rain/fog, seasons, motion blur) and define train/val/test splits for robustness.
* Implement strong baselines (e.g., SegFormer/FCN) with targeted augmentations/invariances (exposure, color, blur, weather).
* Add domain generalization (e.g., style/appearance diversification) and test-time adaptation (entropy- or consistency-based).
* Evaluate robustness across conditions (mIoU/mAP by domain; calibration), and validate in field tests on roadside scenes.

This thesis focuses on the replication and evaluation of two state-of-the-art approaches for multi-target multi-camera tracking (MTMCT), both of which integrate graph-based representations for associating tracked entities across disjoint views. The selected methods—LaMMOn (Language Model + GNN-based tracking) and graph-based tracklet features—demonstrate promising results in online and offline MTMCT scenarios, respectively, using advanced deep learning and graph similarity techniques.

The student will be expected to reproduce the results presented in the following two papers, carefully re-implementing their pipelines and evaluating their performance under consistent experimental settings:

LaMMOn: Language Model Combined Graph Neural Network for MTMCT in Online Scenarios
Link: https://link.springer.com/article/10.1007/s10994-024-06592-1

Graph-based Tracklet Features for MTMCT
Link: https://ieeexplore.ieee.org/abstract/document/10121650

Both methods target real-world intelligent transportation applications and operate on benchmark datasets such as CityFlow. A critical aspect of the thesis will be the comparative evaluation of these methods under controlled conditions to analyze their strengths, limitations, and practical deployment feasibility.

## Central research questions: ##
1. What are the computational and design trade-offs between language model-based and graph-based approaches to MTMCT?
2. How do the two systems compare in terms of tracking accuracy (e.g., IDF1, HOTA), real-time performance (FPS), and modularity?

## Key components of the thesis: ##
1. Re-implementation of both pipelines, respecting the architecture and methods described in the papers.
2. Setup of evaluation experiments on datasets such as CityFlow, TrackCUIP.
3. Quantitative comparison using tracking metrics (e.g., HOTA, IDF1, MOTA, FPS).
4. Qualitative analysis of system behavior in edge cases or failure scenarios.
5. Optional: Visualization tools for tracking results and error patterns.

This thesis investigates the intersection of federated learning and parameter-efficient fine-tuning of Large Language Models (LLMs), with a focus on Low-Rank Adaptation (LoRA) techniques. The goal is to develop and evaluate strategies for fine-tuning LLMs across decentralized clients, each using their own local LoRA modules tailored to their data distributions.

## The central research questions are: ##
1. How can clients collaboratively fine-tune a shared LLM backbone using local, heterogeneous LoRA modules without sharing raw data?
2. How can we effectively aggregate model updates in the presence of statistical heterogeneity across clients?
3. What are the privacy, communication, and performance trade-offs of federated LoRA-based fine-tuning?

The thesis will build upon recent work that shows it is possible to train LLMs in federated environments by exchanging only adapter weights. This project will extend that idea by exploring heterogeneous LoRA configurations—i.e., each client may choose its own rank, structure, or optimization setup—and study how such diversity affects convergence, model utility, and communication efficiency.

## Key components of the thesis: ##
1. Implementation of federated fine-tuning protocols using LoRA and related adapter methods (e.g., QLoRA, AdaLoRA).
2. Support for heterogeneous client-side LoRA configurations (e.g., different ranks, tasks, or target layers).
3. Evaluation on benchmark NLP tasks (e.g., sentiment analysis, question answering) using common federated learning setups (e.g., LEAF, FedNLP).
4. Investigation of aggregation strategies that accommodate non-iid client updates (e.g., FedAvg with normalization, FedProto, FedAdam).
5. Optional exploration of privacy-preserving extensions (e.g., differential privacy, secure aggregation).

## Expected outcomes: ##
1. A practical system that enables federated, privacy-conscious, and communication-efficient fine-tuning of LLMs.
2. Empirical insights into the effects of heterogeneity in LoRA configurations across clients.
3. A set of recommendations for scalable and modular federated LLM training pipelines.

The tactile interpretation of images requires the transformation of two-dimensional visual information into three-dimensional representations accessible through touch. This conversion presents significant challenges, particularly in the preservation of crucial expressive elements that in visual reality are highlighted through light and shadow.

The project focuses on case studies of pictorial art; for example, works by Caravaggio, where chiaroscuro creates a true 'tactile map' through:
- Anatomies highlighted by shadows (muscles, ribs, veins)
- Movement-defining fabric folds
- Depth of environments
- Textures of objects (fruit, fabrics, metals)

The objective of this thesis is to develop an AI system that overcomes current limitations in the automatic conversion of images into tactile representations, focusing on:
- Recognition of elements that in visual reality are highlighted by light and shadow
- Translation of these elements into appropriate depth maps
- Preservation of key narrative elements
- Optimisation of high relief rendering
The project is being developed in parallel with tactile image production experts who will test the effectiveness of the developed solutions.

The project offers the opportunity to:
- Working on a concrete problem of cultural accessibility
- Developing an innovative image interpretation system
- Collaborating with experts in the field of tactile representations
- Contributing to innovation in art accessibility

Tactile sensing presents a promising opportunity for enhancing the interaction capabilities of today’s robots. DIGIT [1] is a commonly used tactile sensor that enables robots to perceive and respond to physical tactile stimuli.

In situations where the sensor is unavailable or experiment repetitions are costly, the value of a reliable, real-time simulation becomes evident. Such a simulation can effectively estimate sensor outputs for various touch scenarios. Such simulation would offer a good alternative to gathering data in different setups and environments.

While several studies have introduced simulations for the DIGIT sensor, such as TACTO [2] and Taxim [3], they predominantly rely on rigid-body simulations and overlook the crucial soft-body aspect of the sensor's gel tip. This oversight diminishes the accuracy of the simulations and their ability to faithfully replicate real-world tactile interactions.


Goals of the thesis

Collect real image outputs and simulated sensor surface deformation.
Train a machine learning algorithm to generate output images using DIGIT surface deformation mesh.
Perform manipulation/grasping tasks with a real UR5 robot to assess and evaluate the performance of the trained algorithm.

Semantic Table Interpretation (STI) is the task of understanding the concepts in a table. This includes (i) cell-entity linking where objects mentioned in table cells are linked to resources such as Wikipedia, (ii) column-type annotation where the type of objects in a column is identified and (iii) column-property annotation which is about finding the relation between two columns [1]. One approach to perform these three tasks is the use of graph neural networks (GNNs) where the table, its columns, rows and cells are represented as a graph [2]. By training on a large corpus, the GNN performs the three tasks through node and edge classification [3].

The goal in this thesis is to develop a GNN that performs the three tasks jointly and to demonstrate that this joint training leads to more accurate annotations than when performing these tasks in isolation.

As a basis of this master thesis, we already have a method for table-to-graph conversion, a basic GNN for STI and an evaluation pipeline setup which will need to be extended to demonstrate the effect of joint training and its superiority over existing approaches.

Your goals are as follows:
- Understand STI and GNNs
- Extend an GNN (implemented in PyTorch Geometric) for STI
- Conduct experiments to evaluate the effectiveness of the GNN