Bayesian Exploration for Morphology-Action Co-optimization
(University of Cambridge, 2019)
Morphology been shown to be a fundamental aspect of tactile sensing in soft robotics, one that can aid, and indeed enable, complex discrimination tasks. For a robot to change its sensor morphology as well as control appropriately, the parametric search over morphology and control parameters is usually slow and unsuited for real-world applications. We develop a framework based on Bayesian Exploration, to allow a robot to co-optimize both changes in tactile sensing morphology and robot action control, to aid in complex tactile object discrimination tasks. We test the framework by performing object discrimination on a set of eight objects, varying three different physical properties: geometry, surface texture, and stiffness. We integrate a capacitive tactile sensor into a flat end-effector and create three soft silicon-based filters with varying morphological properties. We incorporate the end-effector onto a robotic arm and perform repetitive, parameterized touch experiments, on each object. We show morphing is indeed necessary to dissociate amongst different object properties with the sensor at hand. Moreover, we show the proposed framework can consistently achieve optimal morphology-action configurations in approximately half the time than systematic search over parameters. This work marks a step towards the creation of robots capable of using morphology and action control to actively aid in discrimination tasks.