This paper presents a streamlined pipeline for designing, fabricating, and modeling soft robotic swimmers. By combining multimaterial 3D printing with high-performance Peano-HASEL actuators and a differentiable simulation, the researchers successfully bridged the "sim-to-real" gap, allowing a digital model to accurately predict the complex deformations of a physical soft robot.
TL;DR
Researchers at ETH Zurich have developed a 3D-printed soft robotic fish powered by electrostatic "muscles" (Peano-HASELs). By using a differentiable simulator, they can automatically tune the robot's digital twin to match real-world behavior, paving the way for rapid, automated optimization of soft underwater explorers.
What Problem Does This Paper Address?
Soft robots are ideal for ocean exploration because they are compliant, resilient, and mimic the efficiency of biological swimmers. However, they are notoriously difficult to design. Unlike rigid robots, soft bodies undergo complex, non-linear deformations that are hard to model.
Traditionally, engineers rely on "trial and error"—building a robot, testing it, and manually tweaking the design. This is slow and expensive. While simulations exist, there is often a significant "sim-to-real gap," where the digital model fails to capture the nuances of the physical materials and actuators. This paper tackles the challenge of creating a fast, accurate, and differentiable pipeline that connects automated fabrication with high-fidelity simulation.
Methods
The authors employ a three-pronged approach:
- Multimaterial 3D Printing: The swimmer's body is printed using a tool-changing FDM printer. It uses rigid polypropylene (PP) for the "spine" and soft styrene-ethylene-butylene-styrene (SEBS) for the flexible body. This allows for complex, bio-inspired geometries in a single manufacturing step.
- Peano-HASEL Actuators: These are "Hydraulically Amplified Self-healing ELectrostatic" actuators. They consist of oil-filled pouches that contract when a high voltage is applied, mimicking biological muscle fibers. They are lightweight, fast, and efficient.
- Differentiable Simulation (DiffPD): The core innovation is the use of a differentiable Finite Element Method (FEM) simulator. Unlike standard simulations, a differentiable one allows the system to calculate gradients. This means the computer can "see" how changing a parameter (like muscle strength) affects the outcome (swimming speed) and automatically optimize it using calculus-based methods like the Adam optimizer.
Datasets
The study did not use a massive external dataset; instead, the researchers generated a targeted experimental dataset to perform System Identification:
- Training Data: 9 data points consisting of video recordings of the real swimmer at varying voltages (3.5, 4.5, 5.5 kV) and frequencies (0.5, 2.0, 3.5 Hz).
- Validation Data: Unseen test cases at 2.5 Hz across three different voltages (4.25, 4.8, 5.2 kV) to verify the model's generalizability.
- Tracking: The CSRT (Channel and Spatial Reliability Tracking) algorithm was used to extract bending angles from the video footage.
Key Insights and Ideas
- Simplification is Key: Instead of trying to model every microscopic fold of the HASEL actuator, the authors used a simplified "Muscle Model" within the simulation. They modeled the fish as a filled body with contractile elements, then used the differentiable nature of the simulator to "fit" this simple model to the complex reality.
- Canola Oil as a Medium: The fish swims in canola oil rather than water. This serves two purposes: it acts as an electrical insulator for the high-voltage actuators and provides a lower Reynolds number environment, making the fluid dynamics easier to simulate accurately.
- Gradient-Based Sim-to-Real: By treating the simulation parameters as variables in an optimization problem, the researchers could minimize the Mean Squared Error (MSE) between the simulated and real bending angles. This turned "tuning the model" into a standard machine learning task.
Key Results
- High Accuracy: After optimization, the mean absolute angle error between the simulation and the real robot was remarkably low: $1.553 \times 10^{-2}$ radians for training data and $3.103 \times 10^{-2}$ radians for unseen test data.
- Locomotion Optimum: The team identified a clear performance peak; the swimmer reached its maximum forward speed at an actuation frequency of 7 Hz across all tested voltages.
- Force Production: The Peano-HASEL actuators were strong enough to deform the 3D-printed body effectively, though they performed slightly below theoretical analytical models due to "edge effects" in the small-scale pouches.
Limitations and Future Work
- Tethered Operation: The current version is tethered to high-voltage power supplies. Future iterations aim for an untethered, autonomous version with onboard electronics.
- Damping Discrepancies: The authors noted that the simulator (DiffPD) doesn't perfectly capture the damping (energy loss) of the real materials, which affects high-frequency accuracy.
- Fluid-Solid Coupling: The current simulation focuses on the deformation of the body. The next step is to integrate a "learned hydrodynamics model" to simulate how the fish interacts with the surrounding fluid, allowing for full end-to-end shape and control optimization.
Summary updated