TL;DR
This paper introduces a novel method for estimating the relative positions and orientations of two adjacent robotic fish using an artificial lateral line system (ALLS) that mimics fish's natural flow sensing. By applying machine learning techniques to pressure sensor data, the system can accurately determine how one fish is positioned and oriented relative to another, a crucial step for coordinated underwater robot behavior.
What Problem Does This Paper Address?
Coordinated movement and formation control are essential for underwater robot groups, but accurately sensing the relative states (position, orientation, etc.) between individual robots is a significant challenge. Fish, however, excel at this using their natural lateral line system, which detects subtle changes in water flow. This paper aims to replicate this capability by developing an artificial lateral line system (ALLS) for robotic fish that can estimate their relative states, specifically focusing on a leader-follower formation. Previous work has focused on individual robot applications or simpler flow stimuli; this research tackles the more complex scenario of inter-robot sensing using realistic flow patterns generated by one robot's movement.
Methods
The core of this work involves establishing a regression model that links the hydrodynamic pressure variations (HPVs) measured by an ALLS on a "follower" robotic fish to the relative states between it and a "leader" robotic fish.
- Artificial Lateral Line System (ALLS): The system uses an array of nine pressure sensors strategically placed on the follower robot to capture HPVs generated by the leader's oscillating caudal fin.
- Sensor Sensitivity and Redundancy Analysis: Before regression, two novel criteria (Ck1 and Ck2) were developed to quantify the sensitivity of each pressure sensor to variations in relative states. This analysis helps identify which sensors are most informative and determine the optimal number of sensors needed, reducing redundancy and computational load.
- Regression Models: Four common regression algorithms were employed to build the models:
- Random Forest (RF)
- Support Vector Regression (SVR)
- Back Propagation Neural Network (BPNN)
- Multiple Linear Regression (REG)
- Model Evaluation: The performance of these models was compared using Mean Absolute Error (MAE) and the coefficient of determination (R²).
Datasets
The study utilized experimental data collected from flume experiments involving two adjacent robotic fish. The upstream (leader) fish had an oscillating caudal fin, generating flow stimuli. The downstream (follower) fish, equipped with the ALLS, recorded the resulting HPVs. The relative states investigated included distance, oscillating frequency, amplitude, offset, yaw angle, pitch angle, and roll angle. The paper does not explicitly mention publicly available datasets but describes the creation of a "sample set O" for training and validation, with specific sizes detailed for different relative states (e.g., 8750 samples for distance 'd').
Key Insights and Ideas
- Biomimicry for Inter-Robot Sensing: The paper demonstrates that mimicking fish's lateral line system can be highly effective for underwater robots to sense their neighbors, opening doors for more sophisticated multi-robot coordination.
- Sensor Optimization is Crucial: The proposed sensitivity and redundancy analysis (Ck1, Ck2) is a significant contribution, showing that not all sensors are equally valuable and that using fewer, more informative sensors can improve efficiency without sacrificing accuracy.
- Random Forest Outperforms: The Random Forest algorithm consistently showed superior performance in estimating relative states compared to SVR, BPNN, and REG, highlighting its robustness and effectiveness for this type of complex, noisy data.
- Flow Patterns Matter: The use of a realistic flow stimulus (reverse Kármán vortex street from an oscillating caudal fin) is more challenging but more representative of natural underwater interactions than simpler flow conditions.
Key Results
- High Estimation Accuracy: The Random Forest model achieved excellent estimation performance, with R² values generally above 0.95 for most relative states. For example, for relative distance 'd', the best model achieved R² = 0.972 and MAE = 3.250 mm using only 4 pressure sensors.
- Optimal Sensor Count Identified: The analysis determined optimal numbers of pressure sensors (M_r) for each relative state, ranging from 1 (for oscillating amplitude 'A') to 7 (for yaw angle 'α'), demonstrating the effectiveness of the sensor selection criteria.
- Yaw Angle Estimation Challenges: While generally successful, the estimation of relative yaw angle showed degraded performance for larger angles, attributed to a low signal-to-noise ratio in the pressure sensor readings.
Limitations and Future Work
- Hardware Limitations: The paper acknowledges that the hardware deficiency of the pressure sensors and background noise limit estimation accuracy, particularly for larger relative yaw angles.
- Controlled Environment: The experiments were conducted in a flume, a controlled environment. Real-world underwater scenarios with more complex and dynamic flow conditions will present greater challenges.
- Future Work: The authors plan to extend this work to online estimation of relative states for two adjacent freely-swimming robotic fish. They also aim to investigate the impact of the fish's own body motions on the ALLS readings and explore data fusion techniques (e.g., combining ALLS and IMU data) to improve estimation accuracy.
Summary updated