Durations other than 1.0 seem to break learning

[BenBlumer]

The learning seems to get butchered for any duration other than 1.

For example:

if __name__ == '__main__':

  import pylab as plt
  from plot_tools import plot_pos_vel_acc_trajectory

  # only Transformation system (f=0)
  dmp = DiscreteDMP()
  end_time = 1 # Definitely don't change this. 
  frequency = 1000 # Changing this also seems to break things.


  trajectory_time_points = np.linspace(0, end_time, end_time * frequency)
  print "Time begins with %d and ends with %d" %(trajectory_time_points[0], trajectory_time_points[-1])
  trajectory_y_values = [np.sin(10*t) for t in trajectory_time_points]
  dmp.setup(trajectory_y_values[0], trajectory_y_values[-1], end_time)
  dmp.learn_batch(trajectory_y_values, frequency)

  traj = []
  for x in range(end_time * frequency):
    #if x == 500:
    #  dmp.goal = 4.0
    dmp.run_step()
    traj.append([dmp.x, dmp.xd, dmp.xdd])

  fig = plt.figure('f=0 (transformation system only)', figsize=(10, 3))
  ax1 = fig.add_subplot(131)
  ax2 = fig.add_subplot(132)
  ax3 = fig.add_subplot(133)
  plot_pos_vel_acc_trajectory((ax1, ax2, ax3), traj, dmp.delta_t, label='DMP $f=0$', linewidth=1)

  fig.tight_layout()

  plt.show()

Produces a nice sin trajectory, but changing " end_time = 1" to "end_time = 2" produces this:
.

[BenBlumer]

I can get a better trajectory by changing line 180 to "dt = duration / n_sampled" and manually setting tau before learning "dmp.tau = end_time ". But the trajectory still doesn't reach the amplitude that I suspect it should.

Any thoughts?

[carlos22]

I think i discovered the same once. The solution was to transform the trajectory to fit between 0 and 1 and back again after learning. You may find something in my masters thesis [1], but definitly in some of the cited litrature. Good Luck!

[1] http://amser.hs-weingarten.de/cms/administrator/components/com_intranet/uploads/paper/134/Masterthesis.pdf