A lag-compensated kinematic integration network for short-horizon AUV trajectory prediction under degraded navigation observations

Reliable short-horizon trajectory prediction for autonomous underwater vehicles is difficult during Doppler velocity log (DVL) outages because the latest reliable navigation state becomes stale relative to the current vehicle state. This paper presents a lag-compensated kinematic integration network (LC-KIN) that treats this staleness as a displacement of the prediction boundary. A validity mask identifies the latest reliable anchor and its lag; the anchor attitude and body-frame velocity are then used to propagate the state to the forecast origin. Rigid-body kinematics are embedded in the forward path to propagate future pose, while bounded learned corrections capture residual motion beyond first-order propagation. On moderate-degradation (MD-R) and severe-degradation (SD-R) datasets derived from 300 REMUS-100 trajectories, with 45 complete test trajectories treated as independent statistical units, LC-KIN achieves Bridge/Future root-mean-square errors (RMSEs) of 0.695/0.191 m and 3.118/0.320 m, respectively. Compared with the best-performing data-driven sequence baseline, GRU, LC-KIN reduces Bridge/Future RMSE by 46.1%/84.6% on MD-R and 35.7%/93.4% on SD-R. Relative to Diffusion-NC, the Bridge RMSE point estimates decrease by 8.9% and 9.0%, while Future RMSE decreases by 74.6% and 90.1%. On 11 A-KIT field missions, LC-KIN yields the lowest RMSE point estimates among TCN, CTRV, and CV over 2–8 s. These results support lag-aware prediction with explicit kinematic structure under the evaluated degradation conditions and forecast ranges. • A frame-level mask enforces strict input availability during DVL outages. • Lag-compensated re-anchoring reconstructs the forecast origin from a stale anchor. • Bounded corrections and rigid-body integration preserve kinematic structure. • LC-KIN reduces Bridge RMSE by 42.2%/33.8% versus InfoMatchedNet-NC. • Field evaluation extends to 2–8 s; safety assessment is restricted to open-loop analysis.