Hidden Markov model for intelligent extraction of robot trajectory command from demonstrated trajectories

This paper proposes a scheme for selecting the best robot trajectory from a number of demonstrated trajectories. The selection scheme is based on the hidden Markov model (HMM) technique and is divided into four stages. The first stage is the representation of human demonstration by a HMM. The second stage is the preprocessing of input trajectories, which includes transformation of the position trajectory to its frequency spectrum function by the short-time Fourier transformation and mapping of the frequency spectrum function to discretized codes by vector quantization. The third stage relates to the training of the HMM. Having a number of repeated demonstrations, we get multiple observation sequences to tune the HMM parameters so that the trained model is the best one to represent the demonstrations. The last stage is the measurement of the quality of each trajectory. With each trajectory sent through the trained HMM model, a generated likelihood index is obtained which reflects the consistency of the trajectory with the HMM. The trajectory with the maximum likelihood index is considered to be the best for the robot to follow