Computational Sensors for Advancing Soft Robotics

By Soham NandiFeb 27 2024Reviewed by Susha Cheriyedath, M.Sc.

In an article published in the journal Nature, researchers focused on developing compliant strain sensors crucial for soft robots, addressing challenges related to their deformable bodies and dynamic actuation.

The proposed computational sensor design utilized a programmed crack array within a micro-crumples strategy, allowing for highly tunable sensing performance. These sensors demonstrated robust responsiveness under challenging conditions and were integrated into an origami robot with machine intelligence, showcasing potential applications in robotic trajectory prediction and topographical altitude awareness for exploration, rescue operations, and swarming behaviors in complex environments. 

Background

Soft robots, known for their flexibility and agile motion, hold promise for dynamic tasks in unstructured environments, necessitating the integration of compliant strain sensors for intelligent responsiveness. Previous efforts in soft robotics have faced challenges due to the high degree of freedom in robot deformations and unpredictable actuating behaviors. Customizing strain sensors for diverse soft robots often involved time-consuming empirical experiments, hindering predictive design and efficient iterations.

Existing sensor stability tests lacked representation of real-world complexities, as they were mainly conducted under controlled conditions, neglecting external interruptions and dynamic working speeds. This research addressed these challenges through a novel computational strain sensor design. Leveraging a deterministic crack propagation mechanism, the sensors featured laser-programmed interdigital crack arrays within micro-crumples, enabling precise control over crack propagation and tunable sensor characteristics. Finite element analysis (FEA) models based on the sensor structure parameters ensured accurate sensor modeling.

The sensors exhibited remarkable robustness under diverse conditions, including noise interruptions, cyclic loadings, and dynamic frequencies, essential for practical soft robot applications. Integrating these sensors into an origami robot, coupled with machine learning algorithms, enabled autonomous navigation with high accuracy in trajectory prediction and surroundings mapping. This work advanced soft robotics by providing a solution for customizable, predictive, and robust strain sensors, bridging the gap between sensor design complexities and the practical demands of soft robot automation in real-world environments.

Methods

The researchers developed a computational strain sensor design using single-walled carbon nanotubes (SWNT), aiming to address challenges in soft robotic automation. The materials used included SWNT, sodium dodecyl sulfate (SDS), ethanol, silver nanowire, ethyl acetate, polyvinylidene fluoride (PVDF) membrane, and biaxial polystyrene shrink films. The preparation involved creating a SWNT dispersion through ultrasonication, depositing SWNT layers on PVDF membranes, and programming interdigital patterns on these layers using a laser machine.

The strain sensors were fabricated in three different configurations: planar, crumpled, and a specialized sensor named programmed cracks array within micro-crumples (PCAM) sensor. The PCAM sensors exhibited controlled crack propagation behaviors and tunable characteristics, achieved by laser-programmed interdigital crack arrays within micro-crumples. FEA models were employed to accurately simulate the sensing curves based on the sensor structure parameters. Additionally, the sensors demonstrated remarkable robustness under diverse conditions, including noise interruptions, cyclic loadings, and dynamic operation frequencies.

The research also involved the integration of these strain sensors into various soft robotic platforms, such as an origami robot, a pneumatic robot, and a tetrapod microrobot. Machine learning algorithms were applied to a sensor-integrated origami robot for trajectory prediction with less than four percent error and topographical altitude awareness with less than 10% error.

The researchers utilized various experimental techniques, including scanning electron microscope (SEM), atomic force microscope (AFM), optical microscopy, and tensile tests, to characterize the surface morphologies, mechanical properties, and performance of the fabricated sensors. Additionally, a detailed description of the data collection process for training artificial neural network (ANN) models for robotic trajectory and terrain altitude prediction was provided. The ANN models were implemented using the Keras and TensorFlow frameworks. Overall, the researchers presented a comprehensive approach to designing, fabricating, and integrating tunable strain sensors into soft robotic systems, addressing key challenges in soft robotics for enhanced automation in complex environments.

Results

The PCAM sensor is constructed using SWNT on a polystyrene substrate. Its unique design involved laser-programmed crack arrays and thermal shrinkage to create an environmentally stable, piezoresistive strain sensor.
The authors employed computational guidance in the PCAM sensor design, allowing for programmable adjustments in sensitivity and working window. The sensor's structural evolution under strain was explored through FEA and in-situ SEM studies.

Key features of the PCAM sensor included high reproducibility, tunable sensitivity, and a stable linear working window. The sensor demonstrated mechanical stability under various deformations and operated effectively in a dynamic, interrupted mechanical environment.

Furthermore, the PCAM sensor was integrated into soft robots of different scales. The PCAM sensor provided real-time feedback, enabling the robots to navigate and adapt to their surroundings effectively. The research extended to the application of machine learning for trajectory prediction in the origami robot using PCAM sensor data. An intelligent sensor network was developed, showcasing precise trajectory prediction capabilities, outperforming a benchmark model that relied solely on actuation information.

Lastly, the PCAM sensor was utilized in a surrounding awareness scenario, where the integrated robot successfully navigated varied terrains, including climbing hills. An ANN model predicted terrain altitudes with high accuracy based on the PCAM sensor data. 

Conclusion

In conclusion, the authors introduced PCAM sensors for soft robots, employing a computational design with a programmed crack array. Overcoming challenges in soft robotics, these sensors offered tunable and robust performance under diverse conditions. Integrated into an origami robot, they enabled autonomous navigation with high-accuracy trajectory prediction. Bridging the gap between predictive design and practical implementation, PCAM sensors marked a milestone in addressing soft robot complexities. Their adaptability holds promise for applications in confined spaces, unknown terrains, and swarm intelligence scenarios, showcasing advancements in soft robotics and autonomous machine intelligence.