Dynamic liquid volume estimation using optical tactile sensors and spiking neural network

Tactile sensing plays a unique role in robotic applications such as liquid volume estimation. The latest research shows that current computer vision methods for liquid volume estimation are still unsatisfactory in terms of accuracy. Additionally, their performance depends on the environment, such as requiring high illumination or facing issues with occlusions. In the face of these challenges, we design an optical tactile sensor to solve the task of liquid volume estimation. We build the sensor using four photoresistors and a LED on a printed circuit board, and an elastic dome bound with a 3D-printed base by four screw suits. The tactile sensor is attached to the Robotiq and applied in the task of liquid volume estimation in different trajectories movement. The experiment is designed as multiple different trajectories for imitating human movements. We propose the tactile regression spiking neural network (TR-SNN) for the liquid volume estimation. The TR-SNN uses a binary decoding method to decode the output spike trains of the network into liquid volume values. It is a novel and successful spiking neural network (SNN) decoding method for the liquid volume estimation task. The experimental results show that compared to baseline models, the TR-SNN can achieve a state-of-the-art estimation performance with a coefficient of determination up to 0.986.