ceteke / gmm-gmr

GMM+GMR from multiple kinesthetic demonstrations

GMM-GMR

This repository contains the implementation of GMM-GMR based imitation learning from multiple trajectories. Python 2.7 is used for implementation so the code can be ROS compatible.

Installation

$ git clone & cd gmm_gmr
$ python setup.py install

Usage

from gmm_gmr.mixtures import GMM_GMR # Import the GMM_GMR class

... # Load your data and do stuff

gmm_gmr = GMM_GMR(trajectories, n_components)
gmm_gmr.fit()

t, generated_trajectory = gmm_gmr.generate_trajectory(interval)

GMM_GMR(trajectories, n_components)

• trajectories: Trajectories used for training. If these are not aligned i.e list of trajectories, Dynamic Time Warping is used to align the trajectories. If you have aligned trajectories this should be np.ndarray with shape (N,T,D) where N is the number of trajectories, T is the time steps and D is the dimension. Note: If you are using end-effector poses, it is recommended that you convert Quaternions to Euler angles.
• n_components: Number of principle components. This should be less than D.

GMM_GMR.fit()

This function trains PCA and GMM. Returns None.

Warning: The explained variance should be greater than 95% so you can generate nice looking trajectories.

GMM_GMR.generate_trajectory(interval)

This function applies GMR on the time steps that are generated from minimum time step to maximum time step with increases of interval. Then the trajectory is converted from latent space to the original space.

Returns a tuple (time, trajectory time is the time steps and trajectory is the generated trajectory.

Example Output

This is the output of example.py. Given trajectories are end-effector poses with respect to object frame. x axis is time steps and y axis is x,y,z coordinates.