基于可信残差补偿的仿生机器鱼模型预测轨迹跟踪控制

Model Predictive Trajectory Tracking Control for Biomimetic Robotic Fish Based on Confidence-Aware Residual Compensation

摘要: 针对仿生机器鱼因周期性摆尾推进和水动力耦合导致的模型失配问题，提出一种基于可信残差补偿的模型预测控制方法。在低维名义模型基础上，采用时间卷积网络(TCN)以历史轨迹误差和控制输入为特征，预测有限时域内的模型残差；根据预测残差与反馈残差的一致性构造可信系数，对补偿量进行自适应缩放与限幅；并采用DE-NSGA-II离线优化MPC权重参数。MuJoCo仿真表明：与标准MPC相比，所提方法使三工况平均横向误差RMSE降低82.3%，输入变化率降低91.1%；在“8”字形轨迹叠加突变水流扰动下，横向误差RMSE较鲁棒MPC降低47.1%。结果表明，可信残差补偿机制能有效修正模型失配误差，在保证输入平滑性的同时显著提升轨迹跟踪精度。

Abstract: To address model mismatch in biomimetic robotic fish caused by periodic tail-fin propulsion and hydrodynamic coupling, a model predictive control method based on confidence-aware residual compensation is proposed. Based on a low-dimensional nominal model, a temporal convolutional network (TCN) is employed to predict model residuals over a finite prediction horizon using historical tracking errors and control inputs as features. A confidence coefficient is then constructed according to the consistency between the predicted and feedback residuals, enabling adaptive scaling and saturation of the compensation term. In addition, DE-NSGA-II is adopted to optimize the MPC weighting parameters offline. MuJoCo simulation results show that, compared with standard MPC, the proposed method reduces the average lateral tracking-error RMSE by 82.3% and the control-input variation rate by 91.1% across three operating conditions. For a figure-eight trajectory subjected to abrupt current disturbances, the lateral tracking-error RMSE is reduced by 47.1% compared with robust MPC. These results demonstrate that the proposed confidence-aware residual compensation mechanism can effectively mitigate model mismatch, significantly improving trajectory-tracking accuracy while maintaining smooth control inputs.