Bio-to-Robot Transfer of Fish Sensorimotor Dynamics

Real fish EMG, flow tunnel, robotic fish, and flow regimes

Experimental theme: biological muscle signals, hydrodynamic flow regimes, and a robotic-fish counterpart connected by interpretable dynamics.

Overview

Abstract

Swimming in fish arises from integrated neural, muscular, skeletal, and hydrodynamic processes that are hard to capture in compact robotic models. This project presents an interpretable system-identification framework that maps between electromyography (EMG) and kinematics in freely swimming koi, then tests whether the learned dynamics transfer to a robotic fish. A linear ARX architecture captures feedforward EMG-to-tail and feedback tail-to-EMG pathways, extracting interpretable quantities such as delay, gain, natural frequency, and damping. The fish-trained model is then evaluated on processed robotic PWM actuation signals to predict robotic tail displacement, demonstrating direct bio-to-robot transfer without robot-specific retraining.

fish locomotion EMG robotic fish ARX system identification bio-to-robot transfer

Publication

Paper

Published in Advanced Intelligent Systems. The article reports synchronized EMG and kinematic data from freely swimming koi across laminar, Kármán vortex, and reverse Kármán vortex flow regimes, followed by zero-shot evaluation on a robotic fish.

Biological data

Bilateral caudal EMG is rectified, smoothed, differenced left-to-right, and paired with video-derived tail displacement.

Robotic transfer

Processed tail-servo PWM is used as the model input, allowing the fish-trained ARX model to predict robotic tail motion.

Download PDF Open DOI

Method

Interpretable Bio-to-Robot Modeling

A fish-trained ARX model connects biological muscle activity, swimming kinematics, and robotic-fish actuation through a compact interpretable mapping.

1. Signal pairing

Fish EMG envelopes and tail displacement are synchronized into input-output sequences for system identification.

2. ARX dynamics

A low-order ARX model captures temporal input-output structure while preserving interpretable delays and gains.

3. Zero-shot robot test

The same fish-trained model predicts robotic-fish tail displacement from processed PWM signals without retraining.

Signal processing pipeline

Real-fish EMG and robotic PWM are processed into comparable model inputs.

ARX training, cross-validation, residual analysis, and robotic validation workflow

ARX validation workflow

Fish-trained ARX models are evaluated across individuals and transferred to robotic-fish trials.

Results

Cross-Domain Generalization

0.86 ± 0.13

Mean robotic R²

Across 42 robotic datasets, the fish-trained ARX model closely matched measured tail displacement.

97.8%

Higher PFI than DNN

The ARX model substantially outperformed a deep neural-network baseline trained on the same biological datasets.

29 ms

Identified delay

The fitted dynamics expose biologically meaningful timing useful for low-latency robotic control design.

ARX prediction performance and validation metrics on robotic fish datasets

Robotic-fish prediction metrics

Prediction traces and heatmaps show strong agreement across laminar and KVS operating conditions.

Why the interpretable model works

The ARX structure models shared temporal and amplitude relationships: actuation delay, proportional gain, and oscillatory response. That makes it compact, explainable, and robust to the large biological-to-robotic domain shift.

Body-size similarity also matters: groups closer to the robotic fish standard length yielded more faithful transfer, supporting the paper’s morphology-aware interpretation.

Resources

Code and Data

Code repository and related resources will be added here when available. Data supporting the findings are available from the corresponding author upon reasonable request.

BibTeX

Citation

@article{afridi2025bio_to_robot,
  title   = {Bio-to-Robot Transfer of Fish Sensorimotor Dynamics via Interpretable Model},
  author  = {Afridi, Waqar Hussain and Tanveer, Ahsan and Afridi, Rahdar Hussain and Hamza, Muhammad and Wu, Mingxin and Li, Liang and Xie, Guangming},
  journal = {Advanced Intelligent Systems},
  pages   = {e202501117},
  year    = {2025},
  doi     = {10.1002/aisy.202501117}
}