EDAMS: An Encoder-Decoder Architecture for Multi-grasp Soft Sensing Object Recognition

The use of tactile sensing exhibits benefits over visual detection as it can be deployed in occluded environments and can provide deeper information about an object's material properties. Soft hands have increasingly been used for tactile object identification, providing a high degree of adaptability without requiring complex control schemes. In this work, we propose a framework for identifying a range of objects in any pose by exploiting the compliance of a soft hand equipped with distributed tactile sensing. We propose EDAMS, an Encoder-Decoder Architecture for Multi-grasp Soft sensing and an ad-hoc data structure capable of encoding information on multiple grasps, while decoupling the dependency on the pose order. We train the model to map the high-dimensional multi-grasp tactile sensor data into a lower-dimensional latent space capable of achieving the geometrical separation of each object class, and enabling accurate object classification. We provide an empirical analysis of the benefit of multi-grasp perception for object identification, and show its impact on the separation of the objects in sensor space. Notably, we find the classification accuracy to change widely across the number of grasps, ranging from 47.0% for a single grasp, to 99.9% for 10 grasps.