Robotics and Semantic Systems

Computer Science | LTH | Lund University

Robotics and Semantic Systems> Education> Ongoing Master Thesis Projects

Ongoing Master Thesis Projects

Contents

Fit&Flow: Conditional Background and Signal Modelling in High-Energy Physics

Thesis worker: Teo Bergkvist
Supervisor: Eren Aksoy
Examiner: Simon Kristoffersson Lind

Abstract

Particle collisions at the LHC are reconstructed by detectors such as CMS. New particles may appear as small peaks in the resulting data, but identifying them requires an accurate model of ordinary background events. This thesis studies whether machine-learning models that use multidimensional event information can improve signal and background modelling without creating false signals. It also explores the use of normalizing flows for signal modelling.

The research questions are:

  1. Does conditioning the background model on multidimensional event information improve its fit and signal-region prediction compared with context-free sideband fits?
  2. Does a conditional normalizing flow improve signal modelling and extraction compared with a Gaussian or simulation-based template?
  3. Does the flexibility of the conditional background model introduce false signals or bias the estimated signal yield?

The study will use simulated signal events at m = 300 GeV together with a QCD-dominated background dataset. Models will be trained either on the sidebands or on the full invariant-mass distribution, with constraints on model flexibility.

The background will be modelled with a physics-motivated analytical function whose parameters are predicted from per-event context by a multilayer perceptron. This conditional model will be compared with a traditional one-dimensional sideband fit and with the same model without context.

The signal distribution will be modelled with a conditional normalizing flow and compared with a Gaussian or simulation-based template. The signal and background models will then be combined in a likelihood fit to estimate the signal fraction.

Performance will be evaluated using background-yield accuracy in the signal region and signal-injection tests. Background-only tests will be used to check whether the model produces a spurious signal.

Projector-Based Real-Time Marking of Surgical Landmarks for Assisted Visualization in Paediatric Heart Surgery

Thesis workers: Joar Rinaldo Roos, Filip Lennhager
Supervisor: Maj Stenmark
Examiner: Sven Gestegård Robertz

Abstract

The aim of this thesis is to investigate whether a projector-based visualization system can mark points or regions on a surgical object in real time, with sufficient precision to eventually assist paediatric heart surgeons during operations. The project focuses on the combination of video-based analysis, geometric calibration and projection: the system should identify relevant landmarks or regions in a video stream and then project corresponding markings back onto the object.

The proposed research questions are:

  1. Which computer vision and machine learning methods are suitable for identifying relevant surgical landmarks or regions in video data?
  2. How can a camera–projector system be calibrated so that points identified in the image are projected accurately onto a physical object?
  3. What precision, latency and robustness can a prototype achieve, and are the results sufficient to motivate further development towards surgical support?
  4. How can collected video data be used to train models, given that the data contains many different objects in front of the area of interest?

The work begins with a literature review of previous research on surgical visualization, projector-based augmented reality, camera calibration, image registration and machine-learning-based analysis of surgical video. The purpose is to identify suitable methods for detecting relevant landmarks, points or regions in video data and for connecting these to a projector-based visualization system. The scope will be narrowed further depending on which methods and models are available.

AI-Assisted Support: Combining RAG and Codebase Analysis for Automated Resolution Suggestions

Thesis worker: Sixten Rydén
Supervisor: Marcus Klang
Examiner: Pierre Nugues

Abstract

Before support personnel at software companies can respond to a customer or escalate an issue, they spend significant time investigating the incoming ticket: searching through previous cases, reading documentation and locating relevant code. This investigation phase is often repetitive, particularly for recurring issues, and reduces the time available for direct customer interaction.

The goal of this thesis is to design, implement and evaluate an AI system that assists support personnel by automatically suggesting a course of action and a draft customer response when a new support ticket arrives. The system will base its suggestions on two knowledge sources: (1) historical support tickets and their resolutions, and (2) the codebase of the affected product. The vision is to shift the balance of support work away from investigation and towards customer interaction.

The primary research question is: how accurately can a RAG-based system that combines historical ticket retrieval and codebase analysis generate relevant resolution suggestions for incoming support tickets, and how does the choice of knowledge source affect suggestion quality? Secondary questions are how semantic similarity between tickets correlates with the relevance of the suggested resolution, and how the system performs on bug reports compared with general support inquiries.

The system will be built using retrieval-augmented generation (RAG), with two parallel knowledge sources queried depending on the nature of the incoming ticket. Historical support tickets will be chunked, embedded with a sentence-embedding model and stored in a vector database. When a new ticket arrives, it is embedded with the same model and compared with previous tickets. If a sufficiently similar historical case is found, its resolution is used as the basis for a suggested action and customer response, together with a reference to the source ticket.

If no sufficiently similar case exists, the system instead retrieves relevant code from the affected product’s repository using the same vector-similarity principle. A large language model (LLM) then reasons over the retrieved code and suggests a probable cause and resolution path. This output is intended as a suggestion for third-line support rather than as a direct customer response.

A configuration file will control which vector index and code repository the system queries, so that it can be scoped to a specific product. This makes the solution extensible across the company’s entire product portfolio without architectural changes.

Evaluation will be both quantitative, using retrieval metrics such as precision@k and RAGAS-based metrics for faithfulness and answer relevance, and qualitative, through structured feedback from support personnel on the usefulness and accuracy of the suggestions.

The work comprises a literature review of RAG architectures and code retrieval, implementation of the full pipeline, a comparative evaluation of the two knowledge sources separately and in combination, and delivery of a documented system ready for further integration.

Exploring the Trade-Off Surface of Local vs. Cloud LLMs for Confidential BI

Thesis workers: Axel Svensson, Mikolaj Sinicka
Supervisor: Volker Krueger
Examiner: Xuan-Son Vu

Abstract

To keep confidential client data in-house, a small or medium-sized enterprise (SME) must serve its business intelligence (BI) workloads – here, invoice extraction and a BI chatbot – using local open-weight models on limited hardware rather than cloud APIs. The practical question is which local configuration is optimal under these constraints, and how much quality is sacrificed along the dimensions of interest. A configuration involves many interacting levers (e.g. model, size, quantization, serving engine, context length and batching), whose joint impact on accuracy, latency, throughput, cost and CO₂ must be studied with a rigorous, reusable methodology. Cost and CO₂ in particular are rarely combined with accuracy in prior work and form a core part of the novelty. The thesis will map a local-vs.-cloud trade-off surface and derive an evidence-based recommendation from it.

The research questions are:

  1. Where does the local trade-off surface (across model, size, quantization, engine, context and batching) lie relative to a cloud frontier baseline, and what is the Pareto-optimal local configuration under latency and VRAM constraints?
  2. Which local-only model interventions – e.g. distillation or fine-tuning, guided by error attribution (extraction vs. reasoning vs. formatting) – give the most accuracy per unit of latency, cost and energy, and at what training cost?
  3. (Optional) How does a portfolio of task-specialized student models compare with a single generalist model of equal resource budget, once the overhead of routing each request to the right specialist is taken into account?

The experimental setup uses three tiers of variables:

  • Levers (independent variables): model, size, quantization, serving engine, context length, batching, base vs. specialized model, etc.
  • Metrics (dependent variables): accuracy measured with the company’s existing evaluators, latency (TTFT, TPOT and end-to-end), throughput, cost per query, CO₂ per query, VRAM footprint, etc.
  • Scenarios (held fixed, then varied deliberately): query volume and query mix (workload). Scenarios reflect how clients actually use the system and are not parameters to optimize; this ensures that the resulting Pareto front is optimal for representative usage.

For RQ1, a full factorial design is neither feasible nor necessary. In a first pass, accuracy-optimal levers (e.g. model and quantization) are identified and fixed, and the remaining levers are then optimized over the surviving configurations. The cloud frontier model is measured on the same workload and placed on the same axes as a reference point. Local cost per query combines amortized hardware cost (purchase price divided by lifetime query volume) with energy cost (measured energy per query × electricity price). Local CO₂ per query is computed as energy per query × grid carbon intensity. Cloud cost is based on per-token API pricing, and cloud CO₂ is estimated from published provider figures.

For RQ2, a task-specialized student model will be built through distillation and/or fine-tuning. The primary training source is the company’s existing per-query trajectory logs (tool calls and reasoning), supplemented by synthetic query generation for under-represented query types. Two feasibility checks precede the intervention: whether candidate teacher models meet a quality bar (otherwise fine-tuning on human-corrected labels is used instead), and whether tool return values are logged (required for full-trajectory distillation). Error attribution – classifying residual errors as extraction, reasoning or formatting errors – guides the intervention. A student model trained on real logs must remain fully local.

Motion-Based Classification of UAVs and Birds: Generalization to Unseen UAV Platforms

Thesis workers: Alexander Anvin, Edvin Dageson
Supervisor: Simon Kristoffersson Lind
Examiner: Davide Tateo

Abstract

The aim of this project is to investigate how robust motion-based classification of drones and birds is when the model encounters UAV platforms that were not present in the training data.

The project is based on sensor data describing how objects move over time, such as position, velocity and changes in heading. A model is trained on trajectories from several UAV types and from birds, while at least one UAV type is held out entirely from training and used only for testing.

The purpose is to determine whether the model learns general differences between the motion patterns of birds and UAVs, or mainly learns to recognize the platforms it has already seen. The project may also investigate whether the model can recognize when it encounters a motion pattern that deviates from the training data, instead of making a confident but incorrect classification.

The problem is studied at three levels:

  1. Classification: Can the model distinguish between birds and UAVs based on their motion patterns?
  2. Generalization: Does the model also work on a UAV platform or flight behaviour it has not seen before?
  3. Uncertainty: Can the model recognize that it is facing a motion pattern that deviates from the training data, or does it make a confident but incorrect classification?

This leads to the research questions: How well does motion-based classification of drones and birds generalize to previously unseen UAV platforms or flight behaviours? Can the model identify when it encounters motion patterns that deviate from the training data?

The project will use sensor data available at Saab, containing motion information from several UAV platforms and from birds. The data is preprocessed and split so that one or more UAV platforms are held out entirely from the training set.

One or a small number of sequence models, such as LSTM, GRU or Transformer, will then be trained to classify trajectories as UAV or bird. The models are evaluated first on data similar to the training data and then on UAV platforms not seen during training. Comparing performance between these two cases shows how well the models generalize, using metrics such as precision, recall, F1 score and confusion matrices.

As an extension, methods for estimating model uncertainty or detecting out-of-distribution data may be investigated, to see whether the model can signal when it encounters an unfamiliar motion pattern rather than always making a confident classification.

The experiment can be repeated with different UAV platforms as the held-out test group, to give a fairer picture of the models’ ability to generalize.

Last updated: 2026-10-06