Neural ODEs for Data-Driven Automatic Self-Design of Finite-Time Output Feedback Control for Unknown Nonlinear Dynamics

摘要

Many application fields, e.g., robotic surgery, autonomous piloting, and wearable robotics greatly benefit from advances in robotics and automation. A common task is to control an unknown nonlinear system such that its output tracks a desired reference signal for a finite duration of time. A learning control method that automatically and efficiently designs output feedback controllers for this task would greatly boost practicality over time-consuming and labour-intensive manual system identification and controller design methods. In this contribution we propose Automatic Neural Ordinary Differential Equation Control (ANODEC), a data-efficient automatic design of output feedback controllers for finite-time reference tracking in systems with unknown nonlinear dynamics. In a two-step approach, ANODEC first identifies a neural ODE model of the system dynamics from input-output data of the system dynamics and then exploits this data-driven model to learn a neural ODE feedback controller, while requiring no knowledge of the actual system state or its dimensionality. In-silico validation shows that ANODEC is able to —automatically— design competitive controllers that outperform two controller baselines, and achieves an on average ≈ 30% / 17% lower median RMSE. This is demonstrated in four different nonlinear systems using multiple, qualitatively different and even out-of-training-distribution reference signals.