Soft robotic tactile perception of softer objects based on learning of spatiotemporal pressure patterns

The softness perception of objects with lower stiffness than that of robotic skin is challenging, as the proportion of the deformation of skin to that of an object's surface is unknown. This makes it difficult to derive the indentation depth typically used for stiffness estimation. To overcome this challenge, we implemented human-inspired softness sensing in a soft anthropomorphic finger based on tactile information alone without using the information about indentation depth or displacement. In the experiments where LSTM networks were trained to discriminate viscoelastic soft objects, we demonstrated that the sensorized robotic finger using tactile information from barometric sensors embedded in its soft skin could successfully learn to discriminate soft objects. By dissociating the relative contribution of the dynamic pattern of pressure distribution and that of local pressure, we further investigated how differences in available tactile information could impact the ability to distinguish the softness of viscoelastic objects. The results demonstrated that the pressure distribution and its change on the soft contact area of the robotic finger provided information to discriminate the softness of viscoelastic objects and that the tactile information about softness was spatiotemporal in nature. The results further implied that nonlinear local dynamics such as hysteresis in local pressure changes can provide additional information about the viscoelasticity of touched objects.