Background and motivation
Most human activity recognition (HAR) systems using IMU data assume a closed set: all activities at test time are known from training. In real deployments (home monitoring, rehab, industrial settings) people perform unexpected or out‑of‑distribution movements that current models tend to misclassify rather than flag as “unknown,” which is problematic for safety and clinical trust. Recent work in open‑set and zero‑shot HAR suggests that carefully designed representation learning and uncertainty estimation can detect unseen activities while maintaining strong performance on known ones.

Project summary
This project will develop and evaluate methods for open‑set human movement classification using wearable IMU (and optionally EMG) data. The student will build on existing and/or newly collected datasets of functional movements (gait, sit‑to‑stand, reaching, etc.) and design models that (i) classify known activities accurately, and (ii) detect when a test sample is unlike the training distribution and should be treated as “unknown”. Techniques may include conformal prediction, metric learning, distance‑based open‑set thresholds, calibrated confidence scores, and potentially self‑supervised or zero‑shot approaches. The work should aim at a publishable evaluation across several datasets and open‑set protocols, with clear baselines and ablation studies.

Possible aims and research questions

• Can we design features or embeddings from multi‑sensor data that separate known activities from unseen ones?

• What is the impact in performance metrics when open sets are introduced?

• How robust are these methods to sensor placement changes, noisy labels, and mild protocol shifts?

Methods and data

• Possibly use a combination of public HAR datasets (e.g. PAMAP2, MHEALTH, etc.) and internal datasets from your lab (e.g. internship / MSc movement data).

• Implement baseline HAR models (1D CNNs, transformer‑like architectures, or strong classical baselines) and extend them with open‑set mechanisms (e.g. Mahalanobis distance in feature space, OpenMax‑style post‑processing, mixup‑based metric learning).

• Define open‑set evaluation protocols (e.g. leave‑one‑activity‑out, gradual introduction of novel classes) and perform systematic experiments.

• Optionally, explore incremental learning or online adaptation to non‑stationary activity distributions.

Suitability and outcomes

• Suitable for: MSc Computer Science & AI (primary), Data Science, Computer Science.

• Expected outputs: research‑style paper (potentially targeting a HAR / ubiquitous computing / wearable ML venue), open‑source code, and a benchmark comparison for open‑set HAR on several datasets.

Required skills

• Strong Python and ML experience.

• Familiarity with time‑series or sequence models.

Desirable skills

• Background in signal processing or prior exposure to HAR / wearable sensing.

Background and motivation
Machine Learning has significantly improved Human Activity Recognition (HAR) from wearable sensors, but models are often black‑box and hard to trust in clinical or safety‑critical settings. There is growing interest in explainable AI (XAI) for movement analysis to highlight which signal segments or biomechanical features drive a decision. Transparent models could help clinicians, rehabilitation professionals, and end‑users understand misclassifications and assess whether the system is behaving plausibly.

Project summary
This project will develop and evaluate explainable AI methods for IMU‑based human movement classification (e.g. gait types, functional activities, or simple pathology vs. control discrimination). Building on existing movement datasets and or newly collected data, the student will design or adapt models that provide human‑understandable explanations: attention over time, feature‑level contributions (e.g. joint angles, step‑level metrics), or higher‑level movement “concepts”. The project should include both quantitative evaluation (accuracy, stability of explanations) and qualitative analysis (e.g. case studies with domain‑expert feedback if available).

Possible aims and research questions

• Which XAI techniques yield stable, interpretable explanations for IMU‑based HAR?

• Do explanations align with biomechanical expectations (e.g. specific joints) and help identify failure modes?

• Optional: Can we design architectures that are intrinsically more interpretable (e.g. attention over predefined phases or features) without sacrificing performance?

Methods and data

• Use lab datasets (e.g. multi‑IMU movement recordings, potentially with derived kinematics) plus 1–2 public HAR datasets.

• Implement a baseline HAR model (e.g. classical ML, temporal CNN) and attach XAI methods (conformal prediction, saliency maps, temporal attention visualisation, feature attribution).

• Optionally, explore self‑explainable architectures inspired by recent work in explainable HAR and zero‑shot HAR.

• Evaluate explanations quantitatively (e.g. sanity checks, perturbation tests) and qualitatively using simple visual analytics dashboards.

Suitability and outcomes

• Suitable for: MSc Computer Science & AI (primary), Data Science, Computer Science.

• Expected outputs: a research‑style paper on XAI for IMU‑based HAR, code and visualisation tools that other students in the lab can reuse.

Required skills

• Strong Python and machine learning.

• Interest in explainable AI and evaluation of explanations.

Desirable skills

• Some biomechanics or gait analysis background, or willingness to learn basics.

• Experience working with sensor data or time-series.

Background and motivation
As collaborative robots (cobots) increasingly share workspaces with humans, safety and fluent collaboration rely on accurately sensing human motion and intent. Traditional safety systems use fixed guards or simple proximity sensors, which are often too conservative or inflexible for dynamic tasks. Wearable sensors (IMUs, EMG) and human “digital twins” are emerging as powerful tools for real‑time monitoring of worker posture, activity, and intention in industrial and healthcare contexts. Integrating these signals into robot control could improve collision avoidance, shared control, and ergonomic safety.

Project summary
This project will explore how wearable IMU (and possibly EMG) data can be used to infer human motion and intent for safer human–robot collaboration in a shared workspace. The main focus will be on simulation (e.g. simulated cobots / manipulators) plus limited, supervised tests on lab robots if feasible. The student will develop models that: (i) recognise human activities or gestures relevant to collaboration (e.g. approach vs. retreat, reach into shared workspace), and/or (ii) estimate safety‑relevant states (distance, predicted motion) that can modulate robot behaviour (slow‑down, re‑route, stop, yield control).

Possible aims and research questions

• Can wearable‑sensor‑based intent recognition enable smoother and safer shared control policies than purely vision/proximity‑based approaches?

• How can we encode safety envelopes or risk metrics that respond to predicted human motion, not just current position?

• What is a practical interface between wearable‑sensor classifiers and a robot motion planner / controller in simulation?

Methods and data

• Use internal or newly collected IMU datasets for upper‑body and/or whole‑body movement in simple collaborative tasks (e.g. pick‑and‑place on a shared table, handover regions).

• Train ML models for activity/gesture recognition or short‑horizon intent prediction.

• Integrate the classifier into a simulated robot environment (e.g. ROS + Gazebo, or other simulator used in CMS), implementing simple safety behaviours (speed modulation, dynamic avoidance, shared control switching).

• Evaluate performance using standard safety/efficiency metrics (e.g. number of near‑collisions in simulation, task completion time, idle time) and ablation studies.

Suitability and outcomes

• Suitable for: MSc Cyberphysical Systems (primary), Computer Science & AI, Computer Science.

• Expected outputs: research‑style paper on wearable‑sensor‑driven HRC safety/shared control, simulation scenarios and code reusable by other students.

Required skills

• Solid programming in Python and/or C++.

• Interest in robotics, control, or HRC; willingness to work with simulators.

Desirable skills

• Experience with ROS / Gazebo or similar tools.

• Background in ML for time‑series or prior exposure to wearables.

Background and motivation
Wearable sensors and ML are increasingly used to monitor gait, physical activity, and rehabilitation progress in various patient groups (e.g. neurological conditions, musculoskeletal disorders, cardiovascular disease). Systematic reviews highlight promising results, but also methodological gaps: small datasets, limited validation, and models that are not tailored to specific clinical questions. The CHART research group has a number of Health Sciences based PhD students that work with human kinematic and kinetic data, which is a rich opportunity for AI/ML analysis. 

Project summary
This project will work with one or more existing Health Sciences datasets (e.g. IMU‑based movement analysis, rehabilitation tasks, daily‑life activity monitoring) to build and evaluate ML models that address a concrete healthcare question. Examples include: neck pain assessment based on gait and other tasks; sit-to-stand movement analysis for rehabilitation outcome prediction. The emphasis will be on rigorous preprocessing and model evaluation so that results are clinically meaningful and potentially publishable in a healthcare/rehab journal or conference.

Possible aims and research questions

• Which wearable‑derived features (time‑series or biomechanical) are most predictive of the chosen clinical outcome?

• How do different ML models (classical ML vs. deep models) compare under proper cross‑validation and subject‑wise splits?

• Optional: Can we design models that are reasonably interpretable for clinicians (e.g. via feature importance, simple XAI methods)?

Methods and data

• Collaborate with Health Sciences PhD students to understand the clinical question and data collection protocol.

• Implement robust preprocessing: signal cleaning, segmentation, feature extraction, and labelling aligned with clinical annotations.

• Train and evaluate models with careful train/validation/test splits (ideally subject‑wise) and appropriate metrics (AUC, sensitivity/specificity, etc.).

• Optionally, incorporate simple XAI tools to highlight which movement patterns drive predictions, and compare against domain expectations.

Suitability and outcomes

• Suitable for: MSc Data Science (primary), Computer Science & AI, Computer Science.

• Expected outputs: research‑style paper co‑authored with Health Sciences collaborators, plus reusable preprocessing and analysis pipelines.

Required skills

• Strong Python and data‑science stack (Pandas, NumPy, scikit‑learn).

• Comfort working with real, messy data and statistics.

Desirable skills

• Interest in healthcare, rehab, or biomedical applications.