摘要:

Swimming in fish arises from tightly integrated neural, muscular, skeletal, and hydrodynamic processes that are difficult to capture in compact, transferable models for robotics. An interpretable system identification (SySID) is presented that bidirectionally maps between electromyography (EMG) and kinematics in freely swimming koi and further tests its generalization to a robotic fish. Synchronized EMG and kinematic are collected across laminar, Kármán vortex, and reverse Kármán vortex flows spanning 0.146–0.274 m s−1. A linear autoregressive with exogenous input (ARX) model architecture is chosen to capture both feedforward (EMG to kinematics) and feedback (kinematics to EMG) pathways, enabling the extraction of key system parameters, such as natural frequency, damping ratio, and input–output delays. Cross-individual validation demonstrates robust performance and identifies the best-performing fish-trained model, which is then evaluated for cross-domain transfer by replacing EMG input with processed pulse width modulation actuation signals from a robotic fish. Despite differences in mechanics and actuation physics, predictions closely match measured trajectories (mean R2 = 0.86 ± 0.13), substantially outperforming a deep neural network (97.8% higher percentage fit index) trained on the same biological datasets. These findings show that compact, interpretable SySID models enable accurate bio-to-robot transfer without robot-specific retraining, grounding robotic motion models directly in biological function rather than imitation.

Link