Active Bayesian Perception for Simultaneous Object Localization and Identification

In this paper, we propose that active Bayesian per-ception has a general role for Simultaneous Object Localizationand IDentification (SOLID), or deciding where and what. Wetest this claim using a biomimetic fingertip to perceive objectidentity via surface shape at uncertain contact locations. Ourmethod for active Bayesian perception combines decision makingby threshold crossing of the posterior belief with a sensorimotorloop that actively controls sensor location based on those beliefs.Our findings include: (i) active perception with a fixation controlstrategy gives an order-of-magnitude improvement in acuity overpassive perception without sensorimotor feedback; (ii) perceptualacuity improves as the active control requires less belief to makea relocation decision; and (iii) relocation noise further improvesacuity. The best method has aspects that resemble animalperception, supporting wide applicability of these findings.