TOWARDS SEMANTIC TASK EXECUTION ON AERIAL ROBOTS WITH FLOW MATCHING OVER SELF-SUPERVISED VISUAL FEATURES

This thesis studies aerial robot policies built primarily around visual features from DINOv3, a self-supervised vision foundation model, and around flow matching, a generative modeling technique well suited to representing the many equally valid ways a drone can act in a given situation. The same approach is applied to two tasks: reaching a goal position in an unknown indoor environment, and following another drone using vision alone. Supporting both tasks is a GPU-native simulation framework, built by migrating an existing aerial simulator to a new physics backend and extending it with photorealistic indoor scans, color rendering, and distance fields used to generate and evaluate demonstrations at scale. The first task is point-goal navigation. Here we ask whether DINOv3's features serve navigation mainly through the geometric information they encode, or through the broader semantic content the model captures. To the purpose, we evaluate performance of a small adapter under both constrained, hence purely geometric, collision focused, and unconstrained embeddings. The second task is vision-only leader following: the drone must stay behind another vehicle that repeatedly leaves the field of view. The policy learns to segment the leader without ever being supervised to do so. The two tasks lead to different conclusions about memory. For navigation, a simple stateless policy outperforms a recurrent one. The expert knows the whole environment, so its demonstrations never show the kind of behavior that memory would support, such as retracing steps after a wrong turn. For leader following, recurrence is essential, since the target regularly leaves the camera's view and the demonstrations constantly link what was seen earlier to what the drone does next. Arguably, a policy learns to use memory only when its demonstrations require it. The resulting policies generalize to unseen environments without any online correction, and real-world deployment on onboard flight hardware was performed to validate the system.