Event-Domain Trajectory Prediction, Adaptation, and Replanning for USV Formations Under Intermittent Communication

Event-triggered communication reduces unnecessary transmissions in unmanned surface vehicle (USV) formations, but leaves two connected trajectory-level problems. Between leader state packets, the follower side must maintain a leader trajectory prediction; after a new packet arrives, the current and updated reference trajectories may not connect. We propose an event-domain prediction–adaptation–replanning framework to address both problems. Under ideal-link conditions, the leader- and follower-side synchronized event-domain state–residual physics-informed neural network (PINN) trajectory predictors generate identical finite-horizon leader trajectory predictions from identical event histories and online correction states. Event-synchronized residual adaptation updates a low-dimensional online correction state at each event to reduce local prediction mismatch. At each valid reference update, reference trajectory replanning uses a quintic connection to join the current reference trajectory to the updated reference trajectory. In multiseed Circle and S-bend missions, the frozen state–residual PINN predictor reduced short-horizon prediction error relative to nominal dynamics rollout, and the residual-adapted predictor reduced it further. Relative to Direct replacement, quintic connection reduced the mean normalized total control variation by 69.70% across 100 missions.