TL;DR
This paper introduces a novel proprioceptive sensing system for soft underwater tentacles using embedded pressure sensors. By training a Long Short-Term Memory (LSTM) network to reconstruct the tentacle's shape from sensor readings, they enable accurate estimation of swimming performance metrics like tip deflection and traveling wave index, crucial for optimizing control without external vision systems.
What Problem Does This Paper Address?
Soft robots, particularly those designed for underwater locomotion, offer unique advantages due to their compliant nature, enabling complex interactions and efficient propulsion. However, controlling these underactuated, deformable structures is challenging. A key hurdle is understanding the robot's own configuration (proprioception) in real-time, especially in dynamic underwater environments where external sensing like cameras can be unreliable or impractical. Existing methods for estimating soft robot shape often rely on external vision, which can be cumbersome, reduce compliance, or are not suitable for online optimization. This paper aims to bridge this gap by developing an intrinsic sensing method that allows soft tentacles to "feel" their own shape and use this information for performance optimization.
Methods
The core of this work lies in a two-pronged approach: novel sensor integration and advanced model-based reconstruction.
- Sensor Design and Fabrication: Soft tentacles were fabricated using 3D printing, incorporating internal voids at the root. These voids are connected to pressure sensors via flexible tubes. As the tentacle deforms, the volume of these voids changes, altering the pressure, which is then measured. This design ensures the tentacle remains fully soft and compliant. Three pressure sensors were found to provide optimal information content for reconstruction.
- Model-Based Reconstruction: To translate pressure readings into tentacle shape, a Long Short-Term Memory (LSTM) neural network was employed. This network was trained on a dataset correlating pressure sensor readings with ground truth tentacle poses obtained via computer vision. Two models for representing tentacle deformation were compared: a third-degree polynomial and an affine curvature model. The affine curvature model, which describes curvature as a linear function of position along the tentacle's axis, proved more accurate, especially for less stiff tentacles.
- Performance Metrics: The reconstructed tentacle shape is used to derive key performance indicators for swimming:
- Tip Deflection: A measure of how much the tentacle bends at its end, related to propulsion boost.
- Traveling Wave Index (TWI): Quantifies the proportion of deformation driven by a traveling wave versus a standing wave, crucial for efficient locomotion.
Datasets
The paper describes the creation of a training dataset for the LSTM network. This dataset consists of:
- Pressure sensor readings from the embedded sensors.
- Ground truth tentacle poses captured via overhead computer vision.
The tentacle was actuated with randomized triangular wave motions over a range of frequencies and amplitudes. A separate test dataset was also generated for evaluation. The authors also utilized data from experiments on a full robot setup (static and free-swimming) to validate the robustness of their approach.
Key Insights and Ideas
- Intrinsic Sensing for Soft Robots: The paper demonstrates that embedded pressure sensors at the base of a soft tentacle can effectively capture its global deformation, enabling proprioception without compromising softness.
- LSTM for Complex Deformation Mapping: LSTMs are shown to be powerful tools for learning the complex, non-linear relationship between simple pressure readings and intricate tentacle shapes, even with randomized actuation.
- Affine Curvature Model Superiority: For soft, deformable structures like tentacles, the affine curvature model provides a more accurate representation of shape compared to traditional polynomial fitting, particularly for less stiff materials.
- Proprioception for Performance Optimization: The reconstructed shape information (tip deflection, TWI) directly correlates with swimming performance (thrust, velocity), highlighting the potential for proprioceptive feedback to optimize control strategies in real-time, even without external sensors.
- Robustness Across Setups: The trained model successfully transferred from single-tentacle experiments in a tank to a full robot system in a pool, demonstrating the robustness of the sensing and reconstruction approach to different environments and scales.
Key Results
- The affine curvature model achieved a lower normalized root-mean-square error (NRMSE) for tentacle reconstruction compared to polynomial fitting, with segment 2 (closer to the tip) showing up to 2% lower error.
- Absolute tip deflection was predicted with an error of less than 9% in free-swimming tests, without the aid of external sensors.
- In static swimming tests, the reconstructed tip deflection showed a maximum error of only 8.6% compared to ground truth from an overhead camera.
- The full robot achieved a maximum swimming speed of approximately 7 cm/s, with the reconstructed tip deflection correlating well with observed velocity.
- The study successfully demonstrated that proprioceptive sensing can identify optimal actuation parameters for maximizing thrust and efficiency, as indicated by the TWI.
Limitations and Future Work
- Limitations:
- The current model is primarily for planar deformation.
- The study focused on a specific type of actuation (triangular wave).
- The full robot experiments used different servo motors than the single tentacle experiments, potentially introducing variations.
- The paper notes that for the Ecoflex tentacle, polynomial fitting failed to provide accurate tip reconstruction.
- Future Work:
- Extend reconstruction to 3D deformation.
- Optimize sensor design for even greater accuracy and robustness.
- Introduce external objects into the environment to study interaction dynamics.
- Deploy the sensing system on more complex soft structures.
- Further leverage morphological exploitation in underwater environments.
Summary updated