InLaN-Based Dynamics Learning and Task-Space Control for Saturated Spacecraft-Mounted Soft Manipulators

This article establishes the dynamics of spacecraft-mounted soft manipulators (SMSMs) under piecewise constant strain assumption, and then, proposes an Informer-enhanced deep Lagrangian neural network (InLaN) to learn the model of SMSMs. The informer architecture is integrated into the deep Lagrangian neural network, achieving enhanced learning accuracy and extrapolation owing to Informer's ProbSparse self-attention mechanism. In addition, this article designs a finite-time backstepping control strategy for task-space trajectory tracking of SMSMs with the lumped uncertainty being compensated by a high-order sliding mode observer. The range of tracking error is limited by the barrier Lyapunov function, and corresponding compensation is made for input saturation. Simulation results demonstrate that InLaN reduces the 10-s prediction error by over 50% relative to baseline models, and the proposed controller outperforms existing task-space control schemes in terms of tracking performance and robustness.