Feel the touch very important for robotic applications. Here, occlusal phenomena challenge accurate object posture estimation, and subject dynamics are governed by contact interactions.
A recent arXiv.org paper proposes a framework for estimating the posture of an object touched from first touch for objects with known geometry.
Given a 3D model of the object, this method learns a particular object perception model in the simulation. It accounts for the generation of distributions by computing the probabilities between a real tangent shape and a set of simulated tangent shapes. Therefore, additional constraints can be incorporated, such as those arising from multi-contact situations or pre-set object estimations.
Evaluation on real datasets shows that the proposed framework achieves high localization accuracy when exposed to distinct object features.
In this paper, we present Tac2Pose, an object-specific approach to estimate tactile posture from first touch for known objects. Given the object’s geometry, we learn a suitable perceptual model in the simulation to estimate the probability distribution for the possible object poses under tactile observation. To do so, we simulate the contact shapes that a dense set of posing objects would create on the sensor. Then, with a new contact geometry obtained from the sensor, we match it with the precomputed set by embedding the learned specific object using contrast learning. We obtain contact shapes from the sensor with an object-agnostic calibration step that maps RGB tactile observations to binary contact shapes. This mapping, which can be reused across object and sensor instances, is the only step trained with real sensor data. This results in a perceptual model that localizes objects from the first true tactile observation. Importantly, it generates postural distributions and can incorporate additional constraints imposed by other cognitive systems, contacts, or recommender bases.
We provide quantitative results for 20 subjects. Tac2Pose provides highly accurate posture estimates from ad-hoc tactile observations while regressing meaningful postural distributions to explain possible contact shapes as a result of postures. different objects. We also tested Tac2Pose on object models reconstructed from 3D scanners, to assess the degree of certainty to uncertainty in the object model. Finally, we demonstrate the advantages of Tac2Pose over three basic methods for haptic posture estimation: direct object posture regression using a neural network, matching an observed association with a set of possible contacts using a standard classifier neural network and direct pixel comparison of an observed contact with a set of possible contacts.
Website: This http URL
Research articles: Bauza, M., Bronars, A., and Rodriguez, A., “Tac2Pose: Estimating Tactile Object Posture from First Touch”, 2022. Link: https://arxiv.org/abs/2204.11701
