Residual-gated generative additive joint-bias identification and uncertainty-weighted fault compensation for quadruped robots with scarce data

Abstract

Repeated actuator-fault injection on a physical quadruped is costly and can damage motors or gearboxes, which makes continuous fault-value identification difficult when only scarce sequences are available. This study targets command-equivalent additive joint bias. A temporal generator is trained with adversarial and temporal objectives, while an acceptance gate screens synthetic windows using an aggregate score of attitude-integration, first-order joint-kinematic, torque-limit, and command-limit residuals. A heteroscedastic identifier combines dense feature reuse, temporal convolution, bidirectional recurrent features, and cross-attention fusion to estimate the bias and predictive variance. The resulting estimate is converted into an uncertainty-weighted compensation command, with amplitude and rate constraints enforced by final-output projection over their feasible intersection. Controller-level simulations include four additive-bias profiles at the 12 actuated joints and representative same-leg and diagonal two-joint cases. In the reference sinusoidal single-calf-joint evaluation, the method reduces identification RMSE from 0.1404 rad for deterministic DGTNet to 0.1180 rad and reports an interval calibration error of 0.031. With matched controller protections, uncertainty weighting reduces mean recovery time from 0.764 to 0.743 s and RMS command variation from 0.00671 to 0.00651 rad. Across the four single-joint profile groups, the proposed RMSE ranges from 0.1194 to 0.1428 rad, compared with 0.1421 to 0.1696 rad for DGTNet. Same-leg and diagonal paired faults yield RMSE values of 0.1584 and 0.1510 rad and attitude-recovery rates of 74.6% and 77.4%, respectively. Outdoor physical-robot tests with eight runs per configuration show a 54.8% reduction in mean RMS attitude error relative to uncompensated faulty walking, with similar outcomes for constant and uncertainty-weighted confidence.