Search Procedures during Haptic Search in an Unstructured 3D Display Alexandra Moringen 1 , Robert Haschke 2 and Helge Ritter 3 Abstract— In this work, we focus on finding and charac- terizing stereotypical behavioral modes during haptic search in an unstructured 3D display. To this end, we introduce the notion of search procedures and present a machine learning approach for their identification. Search procedures are derived from the haptic exploratory procedures (EPs) through their parameterization and, in some cases, their mixture. In order to identify representative search procedures, we have evaluated data of eight individuals who performed a one-handed haptic search task in nine different scenarios, blindfolded. In all cases, they were asked to localize a specific target object within a tactile scene. Both the target object and the scene were formed from wooden bricks of various tactile shapes arranged in a configurable haptic display to exhibit a specific 3D tactile pattern. By performing t-SNE-based dimensionality reduction and subsequent k-means clustering, we could identify three rep- resentative types of search procedures during manual haptic search: 1) one finger performs one EP 2) several fingers and the hand perform a differently parameterized EPs in a parallel manner 3) a mixture of EPs is performed by the hand and fingers in a parallel manner. We exemplify the notion of search procedures and discuss the corresponding mixtures based on the results of the clustering. I. I NTRODUCTION Haptic search is a process connecting at least three dif- ferent mechanisms in a very efficient way: the movement and the search strategy of the hand and the arm, the haptic and kinematic activity and configuration of the fingers, and the haptic and kinematic perception. In our endeavor to understand this process we can build upon a large amount of previously conducted research in the above-mentioned three directions. Previously conducted modeling of the search behavior of bacteria, fish or foraging animals are commonly used for computational models and robot control (e.g. [12], [10]). Models such as Levy walk, Brownian motion, persis- tent search models and composite correlated random walk seem to yield optimal search strategies depending on the ex- act conditions of the search, i.e. the number, the distribution and type of the targets [8], [7]. Other complex movement strategies, like those of the marching crickets attempting to bite the ones in front and to run away from the approaching ones [1], have not been reused so far. Altogether, it is not yet clear, whether one of the existing models explains the human hand movement strategy during haptic search. Several features, such as velocity and hand configuration, provide us with insights into some aspects of haptic explo- ration. According to Morash [9], hand configuration has an influence on the higher-level strategy of the haptic search. 1,2,3 Neuroinformatics, CITEC, Bielefeld University, Germany {abarch,rhaschke,helge}@techfak.uni-bielefeld.de Fig. 1. Blindfolded participants are presented with two stationary wooden frames containing 10 × 10 (left) and 5 × 5 (right) differently shaped bricks. The task is to memorize the complex shape on the right-hand side and to find it again within the larger frame on the left-hand side, always using a single hand only. In her recent work dedicated to analysis of trajectories in an unstructured display, Morash established a connection between the hand configuration and the choice of exploration strategy. She showed a relation between the detection radius, represented by a hand configuration, and the choice of strat- egy during haptic exploration: systematic vs. non-systematic. On the lower-level, Wing et al. [16] suggest that during haptic exploration, velocity reflects a low level sensorimotor process. Based on the analysis of finger velocity, they showed that different modes can be observed even during haptic exploration of the simplest elliptic shapes with a single finger. A fundamental link between the kinematic and haptic characteristics of the finger and the haptic perception is provided in the work of Hayward [3]. It gives an account of the so called kinematic haptic perceptual invariants, the laws that are invoked in the context of haptic tasks. The author provides an example of a task during which one has to size a coin with a contact movement, and explains which invariants are employed to successfully conduct it [3]. The content of the present paper strongly builds on ex- ploratory procedures, and can be roughly positioned between the higher-level movement strategies and the lower-level sen- sorimotor processes. Our main contributions are as follows: • we introduce the notion of search procedures, typical realizations of EPs during haptic search and • we show that during haptic search EPs are commonly performed together. To clarify the first point, EPs are defined as movement patterns without a particular parameterization. Tanaka at el. [14] demonstrated that the parameterization is modulated by the type of haptic exploration task, discrimination or