A Bio-inspired Multi-functional Tendon-driven Tactile Sensor and Application in Obstacle Avoidance Using Reinforcement Learning

This paper presents a new bio-inspired tactile sensor that is multi-functional and has different sensitivity contact areas. The TacTop area is sensitive and is used for object classification when there is a direct contact. On the other hand, the TacSide area is less sensitive and is used to localize the side contact areas. By connecting tendons from the TacSide area to the TacTop area, the sensor is able to perform multiple detection functions using the same expression region. For the mixed contacting signals collected from the expression region with numerous markers and pins, we build a modified DenseNet121 network which specifically removes all fully connected layers and keeps the rest as a sub-network. The proposed model also contains a global average pooling layer with two branching networks to handle different functions and provide accurate spatial translation of the extracted features. The experimental results demonstrate a high prediction accuracy of 98% for object perception and localization. Furthermore, the new tactile sensor is utilized for obstacle avoidance, where action skills are extracted from human demonstrations and then an action dataset is generated for reinforcement learning to guide robots towards correct responses after contact detection. To evaluate the effectiveness of the proposed framework, several simulations are performed in the MuJoCo environment.