Estimation of Phases for Compliant Motion

Nowadays adding a skill to the robot that can interact with the environment is the primary goal of many researchers. The intelligence of the robot can be achieved by segmenting the manipulation task into phases which are subgoals of the task and identifying the transition between them. This thesis proposes an approach for predicting the number of phases of a compliant motion based manipulation task and estimating their corresponding HMM model that best fit with each segmented phase of the task. Also, it addresses the problem of phase transition monitoring by using recorded data. The captured data is utilized for the building an HMM model, and in the framework of task segmentation, the phase transition addressed. In this thesis, the concept of non-homogeneous HMM is used in modeling the manipulation task, wherein hidden phase depends on observed effect of performing an action (force). The expectation-maximization (EM) algorithm employed in estimating the parameters of the HMM model. The EM algorithm guarantees the estimation of the optimal parameters for each phase of the manipulation task. Hence the modeling accuracy of the forced based transition is significantly enhanced compared to position based transition. To see the performance of the phase transition detection a Viterbi algorithm was implemented. A Cartesian impedance controller defined by [6] for each phase detected is used to reproduce the learned task. The proposed approach is investigated with a KUKA LWR4+ arm in two test setups: in the first, we use parameter estimation for a single demonstration with three phases, and in the second experiment, we find a generalization of the parameter estimation for multiple demonstrations. For both experiments, the transition between phases of the manipulation task is identified. We conclude that our method provides a convenient platform for modeling and estimating of model parameters for phases of manipulation task from single and double demonstrations