Advanced Intelligent Systems · 2026

EMG-Driven Telemetry and Inference System for Fish

Pose Reconstruction and Flow Sensing

Rahdar Hussain Afridi, Waqar Hussain Afridi, Muhammad Hamza, Ahsan Tanveer, Mingxin Wu, Xingwen Zheng, Liang Li, Guangming Xie

Paper Method Results

A unified bio-signal telemetry framework that decodes fish body pose and hydrodynamic context from multichannel intramuscular EMG.

EMG-driven robotic fish pipeline and swimming sequence

Decoded muscle activity can drive a robotic fish in open loop.

16EMG channels

4°joint-angle RMSE

0.98 cmmidline pose RMSE

85.7%generalized classifier accuracy

Bio-signal telemetry and modeling framework

From intramuscular EMG to kinematics and environmental inference

The system records synchronized EMG and video, extracts robust time, frequency, and time-frequency features, augments them with local RMS subsequences, and trains neural models for both regression and classification.

Step 1

Telemetry

Fine-wire electrodes record bilateral axial and fin muscle activity using a custom 16-channel archival unit.

Step 2

Synchronization

Top-view video and EMG are aligned with an LED marker, then fish midlines are extracted with DeepLabCut.

Step 3

Learning

A feature-augmented EMG matrix feeds a deep neural network to predict four body joint angles.

Step 4

Inference

The same EMG pipeline classifies flow speeds and hydrodynamic regimes from muscle activity.

Neural network input-output generation for EMG features and joint angles

Input-output generation: channel features are stacked and mapped to joint angles.

Hardware and experiments

A fish-mounted telemetry unit across structured and unstructured swimming

The study spans free swimming, laminar flow, Karman vortex streets, and reverse Karman vortex streets. This makes the model learn from both naturalistic motion and controlled hydrodynamic perturbations.

EMG electrode layout and telemetry hardware — Electrode layout and telemetry package.

Free-swimming and flow-tank experimental setups — Free-swimming and flow-conditioned trials.

Pose reconstruction

Muscle activity reconstructs whole-body swimming posture

A deep neural network predicts four head-fixed joint angles from 12 axial EMG channels. Forward kinematics then reconstructs the fish midline, enabling downstream estimates of tail displacement, lateral velocity, and energetic output.

Measured and predicted fish joint angles and pose prediction results

Predicted joint angles closely follow measured kinematics across laminar, turbulent, and free-swimming conditions.

Tail displacement, lateral velocity, and total energy output from reconstructed pose

EMG-derived pose supports higher-order swimming metrics such as tail displacement, tail velocity, and total energy output.

EMG preprocessing features and pose extraction pipeline

Raw EMG is filtered, rectified, summarized, and aligned with head-fixed pose variables.

Hydrodynamic condition inference

The same bio-signal encodes flow regime and speed

Replacing the regression head with a softmax classifier turns the pose-decoding pipeline into a flow sensor. Mid-body axial electrodes provide most of the useful information, offering a practical path toward lower-channel telemetry.

EMG channel groups used for flow classification

Channel-group ablations compare axial and fin muscle information.

Flow speed and hydrodynamic regime classification performance

Flow speed and regime confusion matrices show strong diagonal dominance.

Generalized environment and speed classification results

A generalized classifier jointly decodes environment and speed.

Cross-domain validation

Predicted kinematics transfer into CFD and robotic fish motion

The decoded pose is not only a numerical output: it generates hydrodynamic flow fields in simulation and actionable PWM commands for a soft-bodied robotic fish.

CFD validates the hydrodynamic consistency of EMG-derived swimming waves.
Robotic embodiment shows decoded kinematics can become physical locomotor commands.

CFD validation of predicted fish kinematics — CFD vorticity, pressure, and velocity fields.

Open-loop EMG-driven robotic fish actuation — Open-loop robotic fish actuation.

Highlights

Key quantitative findings

Pose regression4 ± 1.4°

mean RMSE for joint-angle prediction.

Pose reconstruction3.84% BL

midline RMSE as a fraction of mean body length.

Flow sensingG3-G5

mid-body axial channels retain most flow information.

General classifier85.73%

best reported accuracy using the full channel group.

Generalized classifier accuracy by channel group

Chart values are taken from the generalized classifier channel-group accuracies reported in the paper.

Citation

BibTeX

@article{afridi2026emgfish,
  title   = {EMG-Driven Telemetry and Inference System for Fish: Pose Reconstruction and Flow Sensing},
  author  = {Afridi, Rahdar Hussain and Afridi, Waqar Hussain and Hamza, Muhammad and Tanveer, Ahsan and Wu, Mingxin and Zheng, Xingwen and Li, Liang and Xie, Guangming},
  journal = {Advanced Intelligent Systems},
  year    = {2026},
  doi     = {10.1002/aisy.202501085}
}