Time-invariant controllers can lead to suboptimal reference tracking for robotic systems faced with variable loads and dynamics. To mitigate this, adaptive control techniques such as model reference adaptive control (MRAC) and adaptive incremental nonlinear dynamic inversion (INDI) adapt to varying dynamics. However, these adaptive control methods require partial knowledge of the system dynamics to obtain the sensitivity derivatives/Jacobian-the direction in which to adapt. The prior knowledge required by these methods can increase the development/design time and limit controller flexibility when faced with highly variable dynamics. Here, inspired by insect control and cerebellum learning systems, we propose an 'implicit' adaptive feedforward (IAFF) control architecture. In contrast to other methods, it does not require any prior knowledge of the system dynamics, and can be deployed in a plug-and-play manner. IAFF uses one filter to learn the sensitivity derivatives and another filter to generate the feedforward actuation command. We demonstrate this adaptive controller in simulations and on a Parrot Mambo mini-drone. We observe that from random initial filter weights, the controller improved the drone's altitude reference tracking capabilities within 30 seconds, and position (pitch and roll) control within minutes. The low computational cost allowed the real-time learning algorithm to be executed on the onboard Arm A9 microprocessor. Copyright (c) 2025 The Authors. This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0/)
原文链接
分享
引用
