Research on Flexible Gait Sensors for Quadruped Robot Applications

Background Introduction

With the widespread application of robots in daily life and industrial production, especially in scenarios requiring standardized, persistent, and heavy-duty operations, the development of intelligent robots has gradually become a trend. However, robots still face many challenges when operating in complex environments, such as rescue missions, automated logistics, autonomous transportation, and smart homes. These robots need to understand their working environments and operate autonomously, with the stability of mechanical movement being a key factor. Traditional methods to ensure stability include using precise sensors to monitor posture and the environment, combined with complex control systems to adjust movements. However, as application scenarios become more complex, existing sensor technologies can no longer meet the demands, particularly in navigating irregular terrains and obstacles.

To address these challenges, researchers have begun exploring new sensor technologies, especially flexible sensors capable of detecting both pressure and vibration. These sensors can mimic the functions of biological mechanoreceptors, helping robots better perceive their external environments. This study proposes a flexible gait sensor based on template-confined electrospinning technology, aiming to provide quadruped robots with high sensitivity and a wide detection range for pressure and vibration.

Source of the Paper

This paper was co-authored by Junyi Ren, Zuqing Yuan, Bin Sun, and Guozhen Shen, affiliated with the College of Physics at Qingdao University, the School of Integrated Circuits and Electronics at Beijing Institute of Technology, and the College of Electronics and Information at Qingdao University. The paper was published on October 24, 2024, in the journal Advanced Fiber Materials.

Research Process

1. Preparation of Electrospinning Solution and Material Selection

The study first prepared the electrospinning solution by dispersing 1.8 grams of thermoplastic polyurethane (TPU) particles in a mixed solvent of 4.1 grams of N,N-dimethylformamide and 4.1 grams of tetrahydrofuran. After magnetic stirring for 5 hours, a uniform TPU solution with a concentration of 18 wt% was obtained.

2. Electrospinning Process and Sensor Fabrication

Researchers used laser engraving to process release paper into a template with densely arranged holes, with diameters of 1.5 mm and 3 mm and a spacing of 0.3 mm between the holes. The template was attached to a metal receiver, and the TPU solution was electrospun through a 23-gauge metal needle at a flow rate of 0.8 mL/h. The electrospinning process was carried out at a voltage of 11 kV, a roller speed of 30 rpm, a temperature of 35±2°C, and a relative humidity of 15%. After completion, the TPU nanofiber membrane with protruding structures was peeled off from the receiver and left at ambient conditions for 12 hours.

3. Sensor Assembly and Encapsulation

Researchers stacked two nanofiber membranes with protruding structures to form a sensitive layer. Subsequently, silver nanowire (AgNWs) dispersion was sprayed onto a polyurethane (PU) encapsulation film to serve as electrodes. Finally, the sensitive layer was sandwiched between two PU films with electrodes and encapsulated using thermal sealing.

4. Material and Sensor Characterization

The morphology of the TPU nanofiber membrane, particularly the cross-section of the protruding structures, was studied using scanning electron microscopy (SEM) and optical microscopy. The capacitance of the sensor was measured using an LCR meter, and vibration signals were collected using a four-channel oscilloscope.

5. Robot Gait Recognition and Deep Learning

Researchers attached the sensors to each leg of a quadruped robot (Xiaomi CyberDog 2) to collect vibration signals under different gaits. By constructing a convolutional neural network (CNN) model, the collected signals were processed for feature extraction and automatic recognition. The training and test sets were divided in an 8:2 ratio, and after 500 training epochs, the model achieved accuracies of 97.50% for gait recognition and 98.04% for abnormal condition detection.

Main Research Findings

1. Pressure Sensing Performance

The study showed that the sensor has a maximum capacitive sensitivity of 1.237 kPa⁻¹, a detection range of up to 1000 kPa, and a response time of 5 ms. Durability tests revealed that the sensor maintained stable performance after 9000 pressure loading cycles.

2. Vibration Sensing Performance

The sensor’s ability to detect vibrations under different weight loads, frequencies, and amplitudes was also validated. Experimental results demonstrated that the sensor could effectively detect vibration signals in the range of 5 Hz to 20 Hz and exhibited good responsiveness to different amplitudes.

3. Gait Recognition

By integrating the sensors with deep learning algorithms, researchers successfully achieved recognition of different gaits of the quadruped robot. Experimental results showed that the sensor could accurately identify gaits such as slow walking, fast walking, running, and bounding, and performed exceptionally well in abnormal condition detection.

Conclusions and Significance

This study developed a flexible gait sensor based on template-confined electrospinning technology, featuring high sensitivity, a wide detection range, and fast response. By integrating deep learning algorithms, the sensor effectively monitored the motion states of quadruped robots and performed well in abnormal condition detection. This research provides new insights into the development of electronic skins for robots and offers potential solutions for enhancing robot performance in complex environments.

Research Highlights

1. High Sensitivity and Wide Detection Range: The sensor demonstrated excellent performance in pressure and vibration detection, with a maximum capacitive sensitivity of 1.237 kPa⁻¹ and a detection range of up to 1000 kPa.
   
2. Fast Response and High Stability: The sensor has a response time of 5 ms and maintained stability after 9000 pressure loading cycles.
   
3. Integration with Deep Learning: Using a convolutional neural network model, the sensor accurately recognized different gaits and abnormal states of the quadruped robot, achieving accuracies of 97.50% and 98.04%, respectively.
   
4. Low-Cost Fabrication: The manufacturing method based on template-confined electrospinning is simple and cost-effective, suitable for large-scale production.

Additional Valuable Information

This study also provided some experimental data and analytical methods, such as analyzing the frequency characteristics of robot motion using Poincaré maps and processing sensor signals through short-time Fourier transform. These methods offer references for future related research.

This research not only has scientific innovation but also provides significant application value for the development of robotics technology.