Enhanced Learning Strategies for Tactile Shape Estimation and Grasp Planning of Unknown Objects

Abstract

Grasping is one of the key capabilities for a robot operating and interacting with humans in a real environment. The conventional approaches require accurate information on both object shape and robotic system modeling. The performance, therefore, can be easily influenced by any noise sensor data or modeling errors. Moreover, identifying the shape of an unknown object under some vision-denied conditions is still a challenging problem in the robotics eld. To address this issue, this thesis investigates the estimation of unknown object shape using tactile exploration and the task-oriented grasp planning for a novel object using enhanced learning techniques. In order to rapidly estimate the shape of an unknown object, this thesis presents a novel multi- fidelity-based optimal sampling method which attempts to improve the existing shape estimation via tactile exploration. Gaussian process regression is used for implicit surface modeling with sequential sampling strategy. The main objective is to make the process of sample point selection more efficient and systematic such that the unknown shape can be estimated fast and accurately with highly limited sample points (e.g., less than 1% of number of data set for the true shape). Specifically, we propose to select the next best sample point based on two optimization criteria: 1) the mutual information (MI) for uncertainty reduction, and 2) the local curvature for fidelity enhancement. The combination of these two objectives leads to an optimal sampling process that balances between the exploration of the whole shape and the exploitation of the local area where the higher fidelity (or more sampling) is required. Simulation and experimental results successfully demonstrate the advantage of the proposed method in terms of estimation speed and accuracy over the conventional one, which allows us to reconstruct recognizable 3D shapes using only around optimally selected 0.4% of the original data set. With the available object shape, this thesis also introduces a knowledge-based approach to quickly generate a task-oriented grasp for a novel object. A comprehensive training dataset which consists of specific tasks and geometrical and physical knowledge of grasping is built up from physical experiment. To analyze and e fficiently utilize the training data, a multi-step clustering algorithm is developed based on a self-organizing map. A number of representative grasps are then selected from the entire training dataset and used to generate a suitable grasp for a novel object. The number of representative grasps is automatically determined using the proposed auto-growing method. In addition, to improve the accuracy and efficiency of the proposed clustering algorithm, we also develop a novel method to localize the initial centroids while capturing the outliers. The results of simulation illustrate that the proposed initialization method and the auto-growing method outperform some conventional approaches in terms of accuracy and efficiency. Furthermore, the proposed knowledge-based grasp planning is also validated on a real robot. The results demonstrate the effectiveness of this approach to generate task-oriented grasps for novel objects.