TL;DR
This paper introduces a novel data-driven approach for controlling flapping-fin Unmanned Underwater Vehicles (UUVs). By developing a search-based inverse model that leverages a neural network predicting thrust from fin kinematics, they enable UUVs to dynamically adjust their fin movements to achieve desired thrust while also optimizing for smooth transitions between movements, crucial for agile underwater navigation.
What Problem Does This Paper Address?
Traditional propeller-driven UUVs, while fast, lack the agility of biological swimmers. Flapping-fin propulsion offers a promising alternative for enhanced maneuverability, especially in complex underwater environments like near-shore areas. However, controlling these fin systems is challenging. Existing methods often rely on pre-determined, high-thrust gaits or incremental adjustments based on limited correlations between fin movements and thrust. This paper addresses the gap in UUV control systems that lack a comprehensive understanding of how fin kinematics influence propulsion, hindering their ability to adapt to dynamic objectives like achieving a target thrust while simultaneously ensuring smooth, stable movements.
Methods
The core of this work is a search-based inverse model that takes desired thrust, the current fin gait, and performance weights as input to output a new, optimized fin gait. This inverse model relies on a forward gait-to-thrust neural network model (specifically, a Long Short-Term Memory or LSTM network) to predict the thrust generated by a given gait.
The inverse model uses a multi-objective loss function that balances thrust accuracy (how close the predicted thrust is to the target) and kinematic smoothness (how much the fin movement changes from one cycle to the next). The weights assigned to these objectives allow the control system to prioritize different aspects of performance on a cycle-by-cycle basis.
Three search-based methods were evaluated for the inverse model:
- Monte Carlo Sampling: Randomly samples potential gaits.
- Hooke-Jeeves Pattern Search (HJPS): A direct search method that explores the solution space.
- Generalized Pattern Search (GPS): An enhanced version of HJPS developed by the authors, which strategically searches for solutions when the algorithm gets stuck in local minima, particularly when kinematic smoothness is heavily weighted.
The forward model uses an LSTM network to predict the time-series thrust output for a flapping cycle, offering a more comprehensive understanding of the gait-thrust relationship compared to simpler models. A dense neural network (DNN) was also evaluated as a baseline but showed higher error.
Datasets
The paper utilizes experimental data collected from a UUV with rigid pectoral fins. This dataset consists of time-series data for five stroke cycles for 864 distinct gaits. These gaits are combinations of static kinematic parameters like stroke amplitude, pitch amplitude, flap frequency, and stroke-pitch offset. The experimental data was used to train and evaluate both the forward and inverse models. For testing the inverse model's ability to handle realistic control scenarios, a simulated UUV position control task using a PID controller was employed, generating 1500 thrust requests.
Key Insights and Ideas
- Dynamic Objective Prioritization: The ability to adjust performance weights (thrust accuracy vs. kinematic smoothness) on a cycle-by-cycle basis is a significant advantage, allowing the UUV to adapt its behavior to the immediate task requirements.
- Search-Based Inverse Models for Real-Time Control: Unlike methods that require retraining for different objectives, search-based inverse models, when coupled with efficient forward models, can adapt their optimization strategy online, making them suitable for real-time control.
- LSTM's Advantage for Time-Series Data: The LSTM network's ability to capture the temporal dynamics of thrust generation proved superior to a simple DNN for predicting thrust, especially for interpolating between unseen gaits.
- Enhanced Pattern Search for Robustness: The proposed GPS algorithm's ability to strategically search upon polling failure makes it more robust, particularly in scenarios where kinematic smoothness is prioritized, preventing it from getting trapped in local minima.
Key Results
- The LSTM forward model achieved a mean average thrust error of 0.0174N on experimental data and 0.0344N when interpolating unseen gaits.
- The proposed GPS inverse model demonstrated the best performance, achieving the lowest mean overall loss of 0.435 and a mean thrust accuracy loss of less than 0.01N when the thrust weight was set to 1.
- All evaluated inverse models (MC, HJPS, GPS) consistently ran within the 0.5-second time constraint required for onboard control on a Raspberry Pi 4.
- The inverse models successfully demonstrated trade-offs between thrust accuracy and kinematic smoothness based on performance metric weights.
Limitations and Future Work
- Negative Thrust Domain: The current experimental data does not cover the domain of negative-thrust gaits. The inverse model currently assumes symmetry for testing, but future work will involve collecting data for negative-thrust gaits and retraining the forward model.
- Propulsive Efficiency: The paper acknowledges that propulsive efficiency is not yet incorporated due to a lack of experimental data for positive flow speeds. Future work aims to extend the inverse model to include this metric, allowing for trade-offs between thrust accuracy and propulsive efficiency.
- Onboard Communication: The final deployment step involves establishing onboard communication between the Raspberry Pi running the inverse model and a PID microcontroller, followed by integration tests with the physical UUV.
Summary updated