Bioinspired Tactile Object Identification Leveraging Deep Learning and Soft Body Compliance

Abstract

Tactile object identification is a fundamental human skill, underlying several core aspects of human intelligence. Humans display a range of remarkable haptic skills, enabled by the synergistic interactions of the somatosensory system with higher-level cognitive processes. In contrast, robotics’ haptic sensing solutions have historically lacked the ability to achieve human-level perceptive capabilities, lacking in both the sensory system and its cognitive digital counterpart. Herein, part of this challenge is addressed by leveraging the success of the fields of soft robotics and deep learning to show how a soft robotic hand, equipped with low-resolution tactile sensing, can be used to accurately identify a diverse set of objects. In particular, ROSE-Net, a neural network that leverages multiple grasps to enable accurate pose-invariant object recognition, is developed. The multi-grasp haptic discrimination solution can lead to a significant increase in performance. The versatility and adaptability of this approach are also tested in two scenarios: a learning transfer scenario and a fault tolerance scenario. Finally, the framework is tested in an online discrimination task, where this approach is shown to naturally require additional grasps for objects that are more challenging to identify using a single grasp and low spatial resolution tactile sensing.