Active Touch Point Selection with Travel Cost in Tactile Exploration for Fast Shape Estimation of Unknown Objects

Tactile sensing for object shape estimation is a key ingredient for real world robots that need to do grasping and manipulation of unknown objects. Since the vision sensors are noisy and suffer from occlusions, the touch to sense approach is beneficial for complementary obtaining local but accurate shape information of objects. Executing exhaustive touches for an object is time consuming and unrealistic, therefore, actively selecting touch points that effectively contribute on the shape estimation is crucial. Prior work suggested to touch the most uncertain part of the estimated shape for minimizing the required number of touches. However, the number of touches may not be the best objective for active touch point selection because it may require an unnecessarily long travel distance and result in an unreasonable execution time. In this paper, we present an alternative approach for active touch point selection by considering the travel costs to the touch points from the current sensor position in addition to the uncertainty of the estimated shape. To estimate the travel costs for all the candidates of touches, we present a graph-based efficient path planning method based on stochastic optimal control. Simulations and real robot experiments with 7DOFs anthropomorphic arm with a single finger device equipped with a tactile sensor are conducted. All the experimental results with several comparisons verify the effectiveness of our method.