Human-Robot Movement Analysis and Learning Using Generative Stochastic Neural Networks

Abstract

This thesis is concerned with movement analysis and learning for human-robot systems which are complex, high-dimensional, and which do not easily admit analytically exact models. Given a set of observations collected from a human-robot system, we address the problems of movement type determination, movement prediction, and trajectory generation, using methods from machine learning. Taking an existing generative stochastic neural network called the conditional restricted Boltzmann machine as our point of departure, we develop a new class of unsupervised learning algorithms for a range of tasks. We first develop an extended version of the conditional restricted Boltzmann machine to predict, in real-time, a lower-limb exoskeleton wearers intended movement type and future trajectory. During training our algorithm automatically clusters unlabeled exoskeletal measurement data into movement types. Our predictor then takes as input a short time series of measurements, and outputs in real-time both the movement type and the forward trajectory time series. Physical experiments with a prototype exoskeleton demonstrate that our method more accurately and stably predicts both movement type and the forward trajectory compared to existing methods. We also apply the developed method to a hand-arm robot for learning reaching and grasping movements. We encode information about the target objects into the stochastic neural network model as observed random variables, through which we can control the system to generate more accurate reaching and grasping movement trajectories in real time. The generated movement trajectories are then fine-adjusted taking into account the contact states between the target object and the hand and fingers so as to ensure good grasp quality. We demonstrate via simulations that our approach can be utilized with a simple control strategy and an existing grasp quality measure to generate appropriate reaching and grasping movements satisfying the force closure condition.