Conclusion and future work

• A closed-loop AFDDR architecture integrating Daphne-AT (diagnosis and decision), PRIDE (procedures), and CRAFTSMAN (robotics) under adjustable human supervision
• habitat-dx: six modular stages; diagnosis and procedure execution are now exercised end to end
• RQ3: multi-channel time-series matching performs best among hand-crafted evaluators (25 correct, 0 wrong, 14 s median)
• Recovery ordering: P/T and highest P both recover 28/31; shortest T and random lose more to the hour limit
• Learned detection: binary LSTM beats threshold (median 3 s vs 7 s); F1 0.92–0.99 across the full noise mode×intensity Monte Carlo grid
• RQ1, RQ2: closed loop enables interaction and autonomy studies; operator and robotics still planned

Future work

• Gate KG / time-series diagnosis from LSTM detections; match Study A schedules
• Hold-out runs / higher-fidelity sims (e.g. BioSim) where train ≠ test generator
• Planning and execution modules sampled from PRIDE and CRAFTSMAN
• Autonomy sweep with operator approvals and iterative robot–diagnosis feedback