Autonomous framework for segmenting robot trajectories of manipulation task

Abstract

In manipulation tasks, motion trajectories are characterized by a set of key phases (i.e., motion primitives). It is therefore important to learn the motion primitives embedded in such tasks from a complete demonstration. In this paper, we propose a core framework that autonomously segments motion trajectories to support the learning of motion primitives. For this purpose, a set of segmentation points is estimated using a Gaussian Mixture Model (GMM) learned after investigating the dimensional subspaces reduced by Principal Component Analysis. The segmentation points can be acquired by two alternative approaches: (1) using a geometrical interpretation of the Gaussians obtained from the learned GMM, and (2) using the weights estimated along the time component of the learned GMM. The main contribution of this paper is the autonomous estimation of the segmentation points based on the GMM learned in a reduced dimensional space. The advantages of such an estimation are as follows: (1) segmentation points without any internal parameters to be manually predefined or pretuned (according to the types of given tasks and/or motion trajectories) can be estimated from a single training data, (2) segmentation points, in which non-linear motion trajectories can be better characterized than by using the original motion trajectories, can be estimated, and (3) natural motion trajectories can be retrieved by temporally rearranging motion segments. The capability of this autonomous segmentation framework is validated by four experiments. In the first experiment, motion segments are evaluated through a comparison with a human expert using a publicly available kitchen dataset. In the second experiment, motion segments are evaluated through a comparison with an existing approach using an open handwriting database. In the third experiment, the segmentation performance is evaluated by retrieving motion trajectories from the reorganization of motion segments. In the fourth experiment, the segmentation performance is evaluated by clustering motion segments.

Unsupervised Trajectory Segmentation and Promoting of Multi-Modal Surgical Demonstrations

“…The range of NMI is [0,1], where 0 means that there is no correlation between two clustering results, while 1 represents the results are completely related. We compare the proposed method TSC-SCAE with stateof-the-art methods, including TSC [8], GMM [7], TSC-VGG, TSC-SIFT [9] and TSC-SCAE on the selected surgical demonstrations. According to the data source in the different methods, the experiments are divided into two categories: one use kinematics data alone and the other use both video and kinematic data.…”

“…1 drawn more attention in recent years. Some unsupervised methods based on Gaussian Mixture Model (GMM) and Dirichlet Processes (DP) are proposed [7], [8]. Although GMM and DP based methods can get rid of the manual annotations, the room to improve the accuracy of surgical trajectory segmentation remains since only the kinematic data is taken into account.…”

A tutorial on task-parameterized movement learning and retrieval

“…Model selection (i.e., determining the number of Gaussians in the GMM) is compatible with the techniques employed in standard GMM, such as the use of a Bayesian information criterion [82], Dirichlet process [22,50,65,74], iterative pairwise replacement [83], spectral clustering [53,69,84] or based on segmentation points [56]. Model selection in mixture modeling shares a similar core challenge as that of data-driven sparse kernel regression techniques, which requires to find the right bandwidth parameters to select a subset of existing/new datapoints that are the most representatives of the dataset.…”

A dynamical system approach to task-adaptation in physical human–robot interaction

“…Although this approach leads to a reliable and smooth transition between tasks, such human-intention recognition strategy (i.e., based on the location of the robot in the workspace) is not efficient; e.g., each task needs a considerable volume of the workspace to be functional, and the robot cannot switch between different tasks in the same area of the workspace. Moreover, there has been recent interesting methods to encode several tasks in one model (Ewerton et al, 2015;Calinon et al, 2014;Lee et al, 2015), and disjointly, several works to recognize and learn the intention of the human (Aarno and Kragic, 2008;Bandyopadhyay et al, 2012;Wang et al, 2018;Ravichandar and Dani, 2015). Only recently, Maeda et al (2017) proposed a probabilistic model that not only encodes for different tasks, but also acts as an inference tool for intention recognition.…”