the real airship in flight · outdoor experiment, FRPG / Perceiving Systems
over the test field · field trialPR-TCN: physics-residual motion prediction for airships
A docking controller needs to know how a blimp will react to a gust before it happens. The prior predictor handled one vehicle in straight flight. I extended it to 17 manoeuvres, five vehicle sizes and four wind models, and replaced the purely learned model with a frozen analytical rollout plus a gust-conditioned TCN that learns only the residual.
The finding I care about most is a negative one: the most accurate analytical prior is not the best one to learn on top of. Of ten architectures across eight held-out shifts, the physics-residual model is the only one that beats the baseline on every shift.
Supervised by Pascal Goldschmid, examined by Jun.-Prof. Dr.-Ing. Aamir Ahmad. Builds on arXiv:2511.19135. Airship photos: Flight Robotics and Perception Group (iFR, University of Stuttgart) / Perceiving Systems (MPI-IS), from Perception-driven Formation Control of Airships (Price, Black, Ahmad) and the Airship-MPC repository. The 1.0 is the top grade on the German scale (A).
















































