Safe autonomous navigation of unmanned aerial systems requires tactical deconfliction to handle conflicts that strategic coordination cannot avoid. Under partial observability, sensor or traffic data may be missing, delayed, or degraded; geometry‑based methods can provide separation guarantees but struggle in dense encounters. Recurrent reinforcement learning has been introduced because it can incorporate observation‑action histories to balance long‑term safety and efficiency.
Operational safeguards such as a remote pilot, onboard detect‑and‑avoid system, or runtime assurance may temporarily override the learned policy. The research gap lies in how recurrent policies behave when authority is handed back after an interruption. Benchmark studies show that interruptions reduce the decodability of task variables from the recurrent hidden state and increase post‑interruption collision risk. The underlying cause is a feedback loop: altered actions change the flight path, which in turn modifies subsequent recurrent updates.
To mitigate this vulnerability, we propose three interruption‑aware architectures that condition recurrent processing on interruption context via separate feature streams, dedicated interruption memory, or gated environmental features. Simulations indicate that all three preserve more decodable task information than a standard recurrent policy. The strongest architecture cuts post‑interruption collisions by 74% and interruption‑induced mission failures by 69% while maintaining interruption‑free performance. These findings demonstrate that a safe interruption does not guarantee a safe handback and support interruption‑aware recurrent design for resilient UAS autonomy. Review