Autonomous active exploration for tactile sensing in robotics

The sense of touch permits humans to directly touch, feel and perceive the state of their surrounding environment. For an exploration task, humans normally reduce uncertainty by actively moving their hands and fingers towards more interesting locations. This active exploration is a sophisticated procedure that involves sensing and perception processes. In robotics, the sense of touch also plays an important role for the development of intelligent systems capable to safely explore and interact with their environment. However, robust and accurate sensing and perception methods, crucial to exploit the benefits offered by the sense of touch, still represents a major research challenge in the field of robotics. A novel method for sensing and perception in robotics using the sense of touch is developed in this research work. This novel active Bayesian perception method, biologically inspired by humans, demonstrates its superiority over passive perception modality, achieving accurate tactile perception with a biomimetic fingertip sensor. The accurate results are accomplished by the accumulation of evidence through the interaction with the environment, and by actively moving the biomimetic fingertip sensor towards better locations to improve perception as humans do. A contour following exploration, commonly used by humans to extract object shape, was used to validate the proposed method using simulated and real objects. The exploration procedure demonstrated the ability of the tactile sensor to autonomously interact, performing active movements to improve the perception from the contour of the objects being explored, in a natural way as humans do. An investigation of the effects on the perception and decisions taken by the combination of the experience acquired along an exploration task with the active Bayesian perception process is also presented. This investigation, based on two novel sensorimotor control strategies (SMC1 and SMC2), was able to improve the performance in speed and accuracy of the exploration task. To exploit the benefits of the control strategies in a realistic exploration, the learning of a forward model and confidence factor was needed. For that reason, a novel method based on the combination of Predicted Information Gain (PIG) and Dynamic Bayesian Networks (DBN) permitted to achieve an online and adaptive learning of the forward model and confidence factor, allowing to improve the performance of the exploration task for both sensorimotor control strategies. Overall, the novel methods presented in this thesis, validated in simulated and real environments, demonstrated to be robust, accurate and suitable for robots to perform autonomous active perception and exploration using the sense touch.