Two complementary VIO pipelines address different failure modes of UAV state estimation: VTIO (thermal + RGB + IMU) is optimised for visual degradation, while RVIO (range + RGB + IMU) is optimised for scale ambiguity and altitude uncertainty. In practice, a drone encounters both failure modes during a single mission, sometimes simultaneously, and a fixed pipeline cannot handle this heterogeneity. In this master thesis, the goal is to develop a meta-system that monitors sensor health in real time and autonomously reconfigures the active estimation pipeline selecting full R-VTIO when all sensors are healthy, VTIO-only when range data is unreliable, RVIO-only when thermal adds no value, or a minimal IMU-propagation fallback during total sensor dropout without introducing state discontinuities at transitions.

The student will define and implement real-time health indicators for each sensor modality: RGB (feature count, spatial distribution, reprojection error), thermal (contrast ratio, feature stability), range (return signal strength, consistency with visual depth), and IMU (vibration spectrum, bias-drift rate). A mode-switching controller will be designed and three strategies will be implemented and compared: rule-based thresholding, a learned classifier (lightweight MLP or decision tree trained on labelled degradation data), and soft blending via dynamic covariance scaling in the factor graph. Seamless state handover between pipeline configurations will be implemented to ensure consistency at switching events. The system will be tested on multi-phase flight scenarios that deliberately trigger multiple degradation types in sequence, benchmarked against fixed-mode baselines, and the work will conclude with a research paper.

The thesis will start with a literature review on multi-modal SLAM and resilient autonomy (MIMOSA, Degradation-Resilient LiDAR-Radar-Inertial Odometry), sensor health metrics, factor-graph optimisation, and existing approaches to mode switching and adaptive covariance.
Next, the student will define the complete experimental framework: definition of per-modality health indicators, implementation of the three switching strategies, integration of seamless state handover between pipeline configurations, and design of multi-phase flight scenarios that exercise the system across multiple degradation transitions.
In the final phase, the student will conduct experimental validation: testing on the Tarot 990 platform across degraded scenarios, comparing switching strategies quantitatively, evaluating tracking-failure reduction against fixed-mode baselines and against manual operator intervention, and analysing failure modes of the autonomous switching layer. The validation phase concludes with a publication-ready research paper.

The ideal candidate has a strong background in robotics, state estimation, and sensor fusion, with proficient C++ and ROS2 skills. Familiarity with factor-graph optimisation (GTSAM, Ceres) is required. Some machine-learning background is helpful for the learned-classifier approach. The candidate should be comfortable working with drone hardware, designing experimental scenarios, and conducting rigorous quantitative benchmarks against ground-truth data.