Data-Driven Control of a Magnetically Actuated Fish-Like Robot

This paper proposes and validates a data-driven control framework for magnetically actuated fish-like robots that combines a neural network-based forward dynamics model with gradient-based model predictive control and imitation learning to achieve precise path following without relying on analytical modeling.

The Gist

Imagine you have a tiny, robotic fish swimming in a bowl of water. This isn't a fish with a battery and a motor in its tail; instead, it's a soft, flexible creature that moves because a giant magnet outside the bowl pulls and pushes it. It's like a puppet, but the puppeteer is an invisible magnetic force.

The problem? Controlling this puppet is incredibly hard. Water is messy and unpredictable, the fish's tail is floppy and doesn't snap back perfectly every time (it has "hysteresis," or memory), and the "commands" you send to the magnet don't last for a fixed amount of time. If you tell the magnet to turn on for 200 milliseconds, the fish moves a little. If you tell it to turn on for 1,000 milliseconds, the fish moves a lot. But because the water fights back differently every time, predicting exactly where the fish will end up is like trying to guess where a leaf will land in a storm.

This paper is about teaching a computer how to control this tricky robotic fish without needing a physics textbook. Here is how they did it, broken down into three simple steps:

1. The "Crystal Ball" (The Forward Dynamics Model)

First, the researchers needed to understand how the fish moves. Instead of trying to write complex math equations about water pressure and magnetic fields (which is nearly impossible for a floppy tail), they just let the fish swim around and recorded what happened.

Think of this as training a Crystal Ball. They showed the computer thousands of examples: "When the fish was here, and we pulled the magnet for this long, it ended up there." They used a Neural Network (a type of AI brain) to memorize these patterns. Now, this AI "Crystal Ball" can look at the fish's current position and say, "If you pull the magnet for 500 milliseconds, the fish will likely end up right here." It learned the rules of the water by watching, not by calculating.

2. The "Chess Grandmaster" (Gradient-Based MPC)

Once the computer had its Crystal Ball, they needed a strategy to make the fish follow a specific path, like a winding river. They used a system called Model Predictive Control (MPC).

Imagine you are playing chess. A grandmaster doesn't just think one move ahead; they think ten moves ahead. They ask, "If I move here, my opponent moves there, then I move there..." to see the best path to victory.

The researchers' computer did the same thing. It used the Crystal Ball to simulate the next 10 steps of the fish's journey. It asked, "If I send this command, where will I be? If I send that one, where will I be?" It kept adjusting its plan until it found the perfect sequence of magnetic pulls to guide the fish along the line. This is the "Grandmaster" strategy. It's very smart, but it's also very slow because it has to do all that thinking in real-time.

3. The "Apprentice" (Imitation Learning)

The "Grandmaster" strategy is too slow for a real-time robot. By the time the computer finishes its 10-step simulation, the fish has already moved, and the plan is outdated.

So, they created an Apprentice. They took the "Grandmaster" (the slow, perfect planner) and let it run thousands of simulations offline. Then, they trained a second, simpler AI (the Apprentice) to watch the Grandmaster and copy its moves.

Think of it like a student watching a master chef. The student doesn't need to understand the chemistry of the ingredients; they just need to learn that "when the sauce looks like this, add a pinch of salt." The Apprentice learned to look at the fish's position and immediately say, "Do this!" without doing the slow, heavy math. It's fast, it's light, and it mimics the perfect plan.

The Result

When they tested this in a computer simulation:

• The Grandmaster (MPC) was able to guide the fish to the target line with incredible precision, missing by less than a centimeter.
• The Apprentice (Imitation Learning) watched the Grandmaster and learned to do the exact same thing, but much faster, with almost the same level of precision.

Why This Matters

This is a big deal because it proves you don't need to be a genius physicist to control a complex, floppy robot in water. You just need to let the robot swim, record the data, and let AI learn the "feel" of the water. This opens the door for tiny, cable-free robots that can swim through coral reefs, inspect pipes, or monitor oceans, guided by a brain that learned to swim just like a real fish.

Technical Summary

Here is a detailed technical summary of the paper "Data-Driven Control of a Magnetically Actuated Fish-Like Robot" by Akiyuki Koyama and Hiroaki Kawashima.

1. Problem Statement

The paper addresses the challenge of precisely controlling magnetically actuated fish-like robots for underwater exploration. While these robots offer advantages in miniaturization and agility (eliminating bulky gears and cables), their control is hindered by three main factors:

• Nonlinear Fluid Dynamics: Unsteady hydrodynamic forces make analytical modeling difficult.
• Flexible Fin Hysteresis: The relationship between magnetic actuation input and the resulting motion of the flexible caudal fin is complex and hysteretic.
• Variable Time Steps: Unlike standard robots with fixed sampling times, the control cycle here is action-dependent. The duration of each control step ($\Delta t_k$) is determined by the coil excitation time (on-time), meaning the physical time step changes dynamically with every action. This complicates standard predictive control models that rely on fixed temporal discretization.

Existing data-driven approaches often rely on conventional servo motors, leaving the specific dynamics of magnetic actuators underexplored.

2. Methodology

The authors propose a comprehensive data-driven control framework that avoids analytical modeling in favor of learning from real-world experimental data. The methodology consists of three core components:

A. Forward Dynamics Model (FDM)

• Architecture: A Multi-Layer Perceptron (MLP) with three fully connected layers and ReLU activation functions.
• Function: It learns the state transition $s_{k+1} = \text{FDM}(s_k, a_k)$.
• Input/Output:
  • State ($s_k$): Defined in a robot-oriented local coordinate system (relative position, orientation, linear velocity, and angular velocity).
  • Action ($a_k$): Consists of the on-times ($b_k, d_k$) for the left and right electromagnetic coils.
• Key Innovation: The FDM is trained to implicitly learn the time-dependency of the dynamics. Since the action variables determine the duration of the step, the network learns to predict the state transition occurring after that specific, variable duration.
• Training: Supervised learning using a dataset of $(s_k, a_k, s_{k+1})$ tuples collected from real-world experiments in a water tank.

B. Gradient-Based Model Predictive Control (G-MPC)

• Role: Acts as an inverse dynamics solver to generate optimal control inputs for path following.
• Mechanism:
  • It optimizes a sequence of control inputs over a prediction horizon $H$ to minimize a cost function $J$ (cumulative error between predicted state and target path).
  • Differentiability: Because the FDM is a neural network, it is differentiable. The G-MPC uses gradient descent (backpropagation) through the FDM to update the action sequence, rather than using black-box optimization methods.
  • Reference Handling: It employs a specific search strategy to select reference points on the target path, balancing position error and heading deviation.
• Limitation: While accurate, the iterative optimization process is computationally expensive for real-time implementation.

C. Imitation Learning Controller (ILC)

• Role: To enable real-time control by approximating the G-MPC policy.
• Mechanism:
  1. Data Generation: The G-MPC is run offline across various scenarios to generate a dataset of "near-optimal" state-action pairs $\{(s_k, p^*_{ref}, a^*_k)\}$.
  2. Training: A separate neural network (MLP) is trained via supervised learning to map the current state and reference point directly to the optimal action.
  3. Inference: During operation, the ILC performs a single forward pass to generate control inputs, drastically reducing computational latency.

3. Key Contributions

1. Variable-Time-Step Modeling: The development of an FDM capable of handling action-dependent time steps, a critical requirement for magnetically actuated systems where control duration varies.
2. Differentiable Control Architecture: The integration of a learned neural network dynamics model into a gradient-based MPC framework, allowing for efficient optimization via backpropagation.
3. Real-Time Approximation: The successful application of imitation learning to replicate the complex G-MPC policy, bridging the gap between high-performance planning and real-time execution.
4. Experimental Validation: The framework is validated using a dataset derived from a physical robot with a flexible fin, addressing the specific challenges of magnetic actuation and fluid dynamics.

4. Results

The approach was evaluated through simulations using the identified dynamics model:

• FDM Accuracy: The neural network successfully captured the nonlinear state transitions.
• G-MPC Performance:
  • The controller achieved accurate path following for a 90-degree turn.
  • RMSE (Root Mean Square Error):
    • Starting above the path: 13.16 mm.
    • Starting below the path: 11.13 mm.
    • Starting on the path: 0.62 mm.
  • The robot converged to the path with minimal oscillation.
• ILC Performance:
  • The imitation learning controller successfully replicated the G-MPC behavior.
  • RMSE: 4.60 mm on the curved path.
  • This demonstrates that the ILC can effectively approximate the optimal policy with significantly lower computational cost.

5. Significance and Future Work

• Significance: This study demonstrates that data-driven strategies are viable for controlling miniature, soft, magnetically actuated robots where traditional analytical modeling fails due to hysteresis and fluid complexity. It provides a pathway for precise navigation in constrained underwater environments.
• Limitations: The current results are based on simulations using the learned model. The physical robustness of the model against real-world disturbances (e.g., water currents, sensor noise) has not yet been fully tested on the physical robot in a closed-loop control setting.
• Future Work: The authors plan to implement the control strategy on the physical robot, starting with simple paths and progressing to complex trajectories. They also aim to investigate the system's robustness against environmental uncertainties and disturbances.