TL;DR
This paper introduces a novel approach to proprioception in soft robots by "multi-tapping" off-the-shelf resistive flex sensors to achieve high-resolution local deformation sensing. Combined with a Finite Element Method (FEM) model, this allows for accurate estimation of both the soft robot's shape and external forces, even in areas not directly covered by the sensor.
What Problem Does This Paper Address?
Soft robots offer significant advantages in human-robot interaction and unstructured environments due to their inherent compliance. However, a major challenge is achieving reliable "proprioception" – the ability of the robot to accurately sense its own shape and external interactions. Traditional methods often rely on simplified assumptions (like constant curvature) or external sensing systems (like motion capture), which are not always practical. Existing embedded sensors often lack the necessary resolution to capture localized deformations, especially when dealing with complex internal fluidic actuation or external forces. This paper aims to bridge this gap by enabling accurate, embedded shape and force estimation in soft robots, crucial for their deployment in real-world scenarios.
Methods
The core innovation lies in modifying standard resistive flex sensors. Instead of using a single resistance measurement, the authors create multiple electrical "taps" along the sensor's resistive layer. This "multi-tapping" allows the sensor to measure resistance changes across multiple segments, providing a much higher resolution of local bending and deformation.
These high-resolution resistance measurements are then fed into a sophisticated Finite Element Method (FEM) model. This model incorporates:
- Solid Mechanics: It captures the physical behavior of the soft robot's material.
- Sensor Model: It translates the multi-tap resistance readings into a shape estimate (pose of the sensor).
- Actuation Model: It accounts for internal forces generated by fluidic actuation (e.g., pneumatic pressure).
The FEM model, built using the SOFA framework, uses these inputs to estimate the robot's overall shape and, crucially, to infer external forces acting on the robot. This is achieved by solving an inverse problem: finding the forces that best explain the observed sensor readings and the robot's known actuation.
Datasets
The paper describes data collection for two primary soft robot prototypes:
- Flexible Strip: A thermoplastic polyurethane (TPU) strip with embedded multi-tap flex sensors. Ground truth for shape was obtained using visual markers tracked by a camera.
- Soft Finger: A pneumatically actuated soft finger designed for a soft gripper. Ground truth for shape was also obtained via visual markers. For force estimation, a robotic arm equipped with a force-torque sensor was used to apply known forces.
The data was collected at 30 Hz and filtered/downsampled. The authors note that unactuated states were under-represented in the initial data, leading to resampling to improve training.
Key Insights and Ideas
- Multi-tapping is key to resolution: Simply using a single flex sensor is insufficient for capturing local deformations. Creating multiple contact points dramatically increases the sensor's ability to resolve fine-grained bending.
- FEM bridges the gap: The FEM model is essential for extrapolating shape estimation beyond the sensor's direct coverage and for integrating actuation and external force information. This is a significant advantage over purely data-driven approaches.
- Proprioception is multi-faceted: The paper emphasizes that true proprioception requires not just shape sensing but also the ability to sense external interactions (forces), and this requires a robust model that accounts for the robot's internal dynamics and actuation.
- Generalization is possible: The FEM-based approach allows the system to estimate shape and forces even in regions not directly instrumented, demonstrating a powerful generalization capability.
Key Results
- Shape Estimation: The multi-tap sensing and FEM modeling approach achieved an average relative error of approximately 3% for shape estimation in the flexible strip and 2.82% for the soft finger, normalized by distance to a reference point. This is a substantial improvement over prior work.
- Force Estimation: For the soft finger, the system could estimate external forces in a range of 0 to 5 N with an average relative error of 11%. This is comparable to vision-based methods and offers a larger force range than some previous data-driven approaches.
- Real-time Performance: The system achieved real-time performance of 30 Hz for the flexible strip and 5 Hz for the more complex soft finger (due to FEM simulation complexity).
Limitations and Future Work
- Nonlinearities and Hysteresis: The authors acknowledge that uncaptured nonlinearities in the soft robot's behavior and hysteresis in the resistive sensors contribute to errors, particularly in force estimation.
- Real-time Optimization: The FEM model's computational cost limits real-time force estimation for complex robots. Optimizing the FEM model for speed (e.g., through model order reduction) is a key future direction.
- Custom Sensor Fabrication: While using off-the-shelf sensors is cost-effective, fabricating custom sensors with integrated multi-tap interfaces could further improve performance and simplify integration.
- Contact Detection: Investigating the possibility of contact detection and force localization solely through shape estimation is another promising avenue.
- Numerical Model in SOFA: Implementing a direct numerical model of the multi-tapped sensor within the SOFA framework could streamline the process of mapping measurements to simulation.
Summary updated