Abstract
Exoskeletons for supporting industrial workers are an emerging technology that have the potential to reduce impact on workers, fatigue and injuries. However its acceptance requires proper and reliable response time. Therefore different technologies will be tested to assess interesting sensor / machine learning algorithms in terms of intent recognition speed and accuracy

Technologies and Approaches
Sensor + ML prediction --> actuator

Hardware
Digital oscilloscope + EMG + IMU + actuator 1 or 2 dimensional

Goals
Provide real time feedback loop

Expected Outcomes
Get an operational feedback loop for the intention prediction, actuate it and measure response times

Contact: Miguel Vazquez Pufleau

Abstract
Exoskeletons for supporting industrial workers are an emerging technology that have the potential to reduce impact on workers, fatigue and injuries. However, its usability should also consider physiological aspects such as the current state of skill of the worker, the degree of expertise and the level o fatigue in order to provide the most adequate response for supporting the movement and the safety of the user.

Technologies and Approaches
Sensor + ML prediction – parameter for support

Hardware
Digital oscilloscope (data acquisition device) + physiological sensor (PH, humidity meter), gyroscope

Goals
Provide real time information on tiredness level / expertise level of user derived from sensor data

Expected Outcomes
Get a functional real time classifier from physiological sensors

Contact: Miguel Vazquez Pufleau

Abstract
The effectiveness of industrial exoskeletons is based on two aspects. First its proper prediction capabilities and secondly the time response. The larger the model, the better the prediction but also the slower the response time. Therefore, a sweet spot needs to be found where the model is good enough but at the same time simple and small enough as to provide the required response time for is intended application

Technologies and Approaches
Sensor + different sizes and types of ML + actuators

Hardware
Digital oscilloscope + intention sensor + computer + actuator

Goals
Develop graphic of the model accuracy vs computational resources vs response time

Expected Outcomes
Assess the effect of different parameters in the model selection on is effect in time critical systems such as an exoskeleton.                 

Contact: Miguel Vazquez Pufleau

Abstract
This project develops models that learn from energy usage patterns and weather data. Using lifelong learning, the system adapts to seasonal changes to schedule home appliances during off-peak hours, reducing electricity costs. It also minimizes grid dependency by optimizing renewable energy use, benefiting smart grids and energy management.

Tasks
Create and test online learning strategies for home appliance scheduling.
Lower energy costs and increase renewable energy usage.
Validate methods with energy forecasting models and deploy prototypes.

Goals
Develop a working prototype that can be tested in real-world energy scheduling scenarios, demonstrating its potential to enhance energy efficiency and reduce reliance on non-renewable resources.                             

Contact: Aftab Hussain

Abstract
This project develops a system to detect and mitigate malicious nodes in Vehicular Ad Hoc Networks (VANETs) using Reinforcement Learning (RL) for real-time decision-making and Continual Learning (CL) to adapt to evolving attack patterns. The approach ensures reliable detection of threats, prevents catastrophic forgetting, and enables collaborative mitigation, enhancing VANET security and resilience.

Tasks
Develop RL-based detection models, integrate CL for adaptability, enable collaborative mitigation via knowledge sharing, and validate the system in simulated and real-world scenarios.

Goals
Build a scalable system for online malicious node detection and mitigation, ensuring adaptability to evolving threats and robust VANET security.

Contact: Aftab Hussain

Abstract
This project focuses on developing methods to detect and classify malicious nodes in IoT networks using Streaming AI. Lightweight, resource-efficient models will monitor network activity and classify nodes in real-time. By leveraging distributed knowledge streams, IoT nodes can collaboratively detect threats and adapt to evolving attack patterns, ensuring robust network security for large-scale IoT systems.

Tasks
Develop and test anomaly detection and node classification strategies for IoT networks.
Enable real-time knowledge sharing across the network to identify and mitigate malicious nodes.
Validate the approach through real-world IoT security use cases.

Goals
Develop a Streaming AI-based model for real-time detection and classification of malicious nodes in IoT networks, leveraging incremental learning and distributed knowledge streams to enhance security and adaptability in dynamic environments.

Contact: Aftab Hussain

Abstract
This thesis looks at using few-shot learning (training AI with very little data) in decentralized systems. It aims to improve how these systems work when data is scarce or imbalanced.

Tasks
Study few-shot learning methods and where they are used.
Design models that can handle small datasets in decentralized systems.
Test how these models perform and their communication costs.

Goals
Make AI systems work better with less data.
Solve the problem of data imbalance in decentralized setups.
Find efficient ways to use few-shot learning in real-world applications.

Key Skills
Few-shot learning basics, understanding decentralized systems, Python.

Contact: Waleed Khan

Abstract
Focus of this thesis is about creating a simple tool to shrink AI models using techniques like pruning (removing less important parts), quantization (simplifying numbers), and sparsification (removing redundant data). This helps run AI on small devices like IoT sensors by reducing the size and communication needs of the models.

Tasks
Learn about existing methods to make AI models smaller.
Create and test these methods on models used in decentralized systems.
Measure how well the smaller models perform and how much they save on communication and energy.

Goals
Make AI models smaller without losing much accuracy.
Save communication resources to make systems scalable.
Improve AI for real-world use on small devices.

Key Skills
Python, TensorFlow Lite/PyTorch, basic research, AI model optimization

Contact: Waleed Khan

Abstract
The thesis focuses on creating a smart way to choose which devices in a network should share AI updates. It focuses on making communication efficient and ensuring the system scales well as the network grows.

Tasks
Review how peer selection works in distributed systems.
Design a method to choose peers based on similarity, speed, or available resources.
Test the method in a simulated federated learning setup.

Goals
Make AI systems communicate more efficiently.
Keep systems running smoothly even in changing environments.
Create a scalable peer selection strategy.

Key Skills
Python, distributed systems basics, creating algorithms.

Contact: Waleed Khan

Abstract
Experts are known to have significantly different gaze-patterns compared to novices. Those gaze-patterns can help to reveal important visual aspects, e.g. in visual search tasks, manufacturing processes, quality control, etc.
Eye-tracking allows to extract these gaze patterns and thus determine those areas of visual importance while augmented reality can be utilized to provide visual clues to novices.

Tasks
Design and implement an experimental framework to extract gaze patterns based on a head-worn eye-tracker and provide visual cues to novices via an augmented reality headset. Conduct a user study to empirically analyse the differences in novice’s performance when visual cues are provided for certain tasks.

Hardware
Pupil Labs Neon head worn eye-tracker. Microsoft HoloLens AR headset.

Goals
Create a prototypical implementation that can be applied to real-world industrial scenarios. Empirically analyse the impact of visual cues on task performance.

Contact: Martin Schobesberger

Abstract:
Development of a human body physics library using real IMU data, enhanced through data augmentation and machine learning (ML) techniques, to account for various body types and industrial activities.

Task

Capture human movement data using a motion capture (MoCap) system (OptiTrack). Develop and train data-driven ML models using a state-of-the-art surface mesh model, such as SMPL. The process must be fully traceable and controllable, ensuring reproducibility. Some required hardware and software will be provided by the department.

Goals

Deliver a validated human body physics library and a fully functional toolchain.

Abstract
This work explores technological solutions for Textile Identity Management that avoid embedding components (e.g., conductive materials, resistors, transducers), tagging (RFID, NFC), and labeling (QR codes), which often complicate recycling. Instead, it proposes utilizing the intrinsic structure of textile fibers to encode identity, ensuring zero additional recycling challenges.

Task
Integrate structure-coded fibers and AI-powered scanners into textile production and recycling, eliminating the need for external tags or labels while addressing material identification challenges in mixed fibers.

Goal
Develop a zero-waste textile identity system that enables accurate material sorting and enhances recyclability.

Abstract
This work investigates technological solutions for Textile Identity Management in compliance with forthcoming EU regulations. The focus is on integrating essential information throughout the entire lifecycle of a textile product into a Digital Product Passport (DPP). A DPP serves as a comprehensive digital identity, continuously updated by all stakeholders involved in the production, use, and recycling of garments.

Task
Develop and integrate relevant APIs for the DPP, contribute to ongoing standardization efforts, and create a prototype that demonstrates textile identity management.

Goal
Design a prototype interface for the DPP that supports an identity management system, enabling precise material sorting and improving recyclability.

Abstract
The project focuses on developing methods and models that can actively learn from dynamic knowledge streams while contributing to them. Knowledge streams, representing continuously evolving data, require adaptive learning approaches to keep up with real-time changes. This is particularly important for industrial applications, where data-driven decision-making must be scalable and responsive. The project aims to leverage reinforcement learning to improve the ability of AI systems to learn from and enhance these streams.

Tasks
Design and implement an experimental framework to explore and develop new learning strategies tailored for knowledge stream-based reinforcement learning. The results will be validated using industrial use cases as part of the Streaming AI initiative.

Goals
Create a scalable, prototypical implementation that can be applied to real-world industrial scenarios.

Abstract
The project aims to develop methods and models capable of running on small, resource-constrained embedded systems. These systems will learn from their environment and contribute to continuous knowledge streams, facilitating updates across other nodes in the network. This approach enables distributed learning and adaptation, essential for real-time applications in industrial environments.

Tasks
Design and implement an experimental setup to explore new learning strategies for embedded devices and generate knowledge streams. Validate these strategies through industrial use cases as part of the Streaming AI initiative.

Goals
Develop a scalable, prototypical implementation that is applicable to real-world industrial scenarios.

Abstract
The project aims to develop methods and models capable of actively learning from their environment through perception. Using relevant sensor input, local embedded systems will continuously perceive and learn about their surroundings, updating their knowledge in real time to enhance decision-making.

Tasks
Design and implement an experimental framework to explore and develop new learning strategies for perception-based learning. Validate these strategies with industrial use cases as part of the Streaming AI initiative.

Goals
Create a scalable, prototypical implementation that can be applied to real-world industrial scenarios.

Abstract
The project aims to develop methods and models for wearable intelligence systems that can actively learn from their environment and user interactions. Equipped with sensors, these wearable devices will continuously gather data, adapt to user behaviour, and update their knowledge to provide personalized, real-time insights and assistance.

Tasks
Create an experimental setup to explore and develop new learning strategies for wearable devices, focusing on context-aware learning and user interaction. Validate these strategies through real-world scenarios as part of the Wearable AI initiative.

Goals
Develop a scalable, prototypical implementation that is applicable to everyday use cases in health, fitness, and personal assistance.

Abstract
The project aims to develop methods and models for body-worn intelligence systems that actively learn from the user’s physical environment and activities. Using embedded sensors, these systems will continuously monitor user movements, adapt to changing conditions, and update their knowledge to provide real-time, context-aware feedback and support.

Tasks
Design and implement an experimental framework to explore new learning strategies for body-worn devices, focusing on activity recognition and environmental awareness. Validate these approaches through practical applications as part of the Body-Worn AI initiative.

Goals
Develop a scalable, prototypical implementation that can be applied to real-world scenarios in health, safety, and performance monitoring.

Abstract
Training and development of a DL model to estimate the elasticity of garments worn on the human body. Imagine wearing a garment with integrated IMU sensors that continuously measure orientation. Existing approaches use only the orientation information to reconstruct the 3D surface shape of the wearer. Problem: rigid reconstruction methods do not consider the material’s elasticity.

Technologies and Approaches
Create and train a (supervised) model that learns from the dense orientation information, attached to surface points, to link it with the expansion attributes of the material to produce more realistic and accurate models. The heuristic approach can potentially be improved with more input parameters, such as material or gender specifics. Go beyond isometric deformations to overcome the limitations of non-stretch surface reconstruction algorithms.

Data material
Real and Synthetic IMU recordings. Methods/Tools: DL meth., VTK, OpenCV

Abstract
Wearing a garment with integrated IMU sensors that continuously measure orientation and estimate the wearer's surface using 3D reconstruction methods. Examine the reconstruction quality and effort (energy,time) in relation to different placements, and different densities for different body regions or activities.

Technologies and Approaches
In analogy to the mechanosensitive organs in human skin (Merkel cells: ~ 60 mils, Ruffini /Meissner corpuscles: ~0.5 mils, Vater-Pacini vibrissae: ~40,000), requirements define densities.
Detailed question: The orientation of an IMU can be expressed by direction vectors with a fixed length, by aggregation we can reconstruct the shape. What if we allowed different vector lengths - equivalent to different distances between sensors?

Data material
Real and Synthetic IMU Recordings.

Methods/Tools
DL meth., VTK, OpenCV

Abstract
One of the biggest issues in hands-free interaction is the feedback to the user whether or not an input has been recognized/accepted by the system. With use of body-worn vibro-tactile devices, such information can be brought to the user to verify the success of an input.

Technologies and Approaches
Computer Vision, eye-tracking, user studies, statistical analysis, questionnaires, interviews, machine learning

Hardware
Vibro-tactile devices (Smartwatch), eye-tracker

Goals
Analysis of the impact on accuracy and understandability of vibro-tactile based feedback compared to other methods (verbal, visual, no feedback, etc.)

Expected Outcomes
Soft-/Hardware prototype, Statistically proven results, written thesis

Abstract
Measuring the mental demand of tasks – also known as Cognitive Load (CL) – has been well researched. However, traditionally such measurements are based on eye-tracking or questionnaires. To make CL more accessible in day-to-day environments, the use of IMUs for on-body motion tracking to determine CL shall be studied.

Technologies and Approaches
Motion tracking, eye-tracking, user studies, statistical analysis, questionnaires, interviews, machine learning

Hardware
IMUs, eye-tracker

Goals
Analysis of the accuracy of motion based CL measurement compared to other methods (eye based, self-perceived, ....)

Expected Outcomes
Soft-/Hardware prototype, statistically proven results, written thesis