Tactile Object Pose Estimation from the First Touch with Geometric\n Contact Rendering

Abstract

In this paper, we present an approach to tactile pose estimation from the\nfirst touch for known objects. First, we create an object-agnostic map from\nreal tactile observations to contact shapes. Next, for a new object with known\ngeometry, we learn a tailored perception model completely in simulation. To do\nso, we simulate the contact shapes that a dense set of object poses would\nproduce on the sensor. Then, given a new contact shape obtained from the sensor\noutput, we match it against the pre-computed set using the object-specific\nembedding learned purely in simulation using contrastive learning.\n This results in a perception model that can localize objects from a single\ntactile observation. It also allows reasoning over pose distributions and\nincluding additional pose constraints coming from other perception systems or\nmultiple contacts. We provide quantitative results for four objects. Our\napproach provides high accuracy pose estimations from distinctive tactile\nobservations while regressing pose distributions to account for those contact\nshapes that could result from different object poses. We further extend and\ntest our approach in multi-contact scenarios where several tactile sensors are\nsimultaneously in contact with the object. Website:\nhttp://mcube.mit.edu/research/tactile_loc_first_touch.html\n