def test_dmp_imitate_pseudoinverse(): x0, g, execution_time, dt = np.zeros(2), np.ones(2), 1.0, 0.01 beh = DMPBehavior(execution_time, dt, 200) beh.init(6, 6) beh.set_meta_parameters(["x0", "g"], [x0, g]) X_demo = make_minimum_jerk(x0, g, execution_time, dt)[0] beh.imitate(X_demo) X = beh.trajectory()[0] assert_array_almost_equal(X_demo.T[0], X, decimal=2)
def test_dmp_save_and_load2(): import pickle with open("reload_test_trajectory_and_dt.pickle", "r") as f: recorded_trajectory, dt = pickle.load(f) artificial_trajectory = np.zeros((500, 1)) artificial_trajectory[:, 0] = np.sin(np.linspace(0, 500, 1)) trajectories = [artificial_trajectory, recorded_trajectory] for trajectory in trajectories: n_steps = trajectory.shape[0] n_task_dims = trajectory.shape[1] execution_time = dt * (n_steps - 1) beh_original = DMPBehavior(execution_time=execution_time, dt=dt, n_features=20) beh_original.init(3 * n_task_dims, 3 * n_task_dims) x0 = trajectory[0, :] g = trajectory[-1, :] beh_original.set_meta_parameters(["x0", "g"], [x0, g]) X = np.zeros((n_task_dims, n_steps, 1)) X[:, :, 0] = np.swapaxes(trajectory, axis1=1, axis2=0) beh_original.imitate(X, alpha=0.01) xva = np.zeros(3 * n_task_dims) xva[:n_task_dims] = x0 beh_original.reset() imitated_trajectory = beh_original.trajectory() try: beh_original.save("tmp_dmp_model.yaml") beh_original.save_config("tmp_dmp_config.yaml") beh_loaded = DMPBehavior(configuration_file="tmp_dmp_model.yaml") beh_loaded.init(3 * n_task_dims, 3 * n_task_dims) beh_loaded.load_config("tmp_dmp_config.yaml") finally: if os.path.exists("tmp_dmp_model.yaml"): os.remove("tmp_dmp_model.yaml") if os.path.exists("tmp_dmp_config.yaml"): os.remove("tmp_dmp_config.yaml") beh_loaded.reset() reimitated_trajectory = beh_loaded.trajectory() assert_array_almost_equal(imitated_trajectory, reimitated_trajectory, decimal=4)
def test_shape_trajectory_imitate(): n_step_evaluations = range(2, 100) for n_steps in n_step_evaluations: n_task_dims = 1 dt = 1.0 / 60 # 60 Hertz execution_time = dt * (n_steps - 1 ) # -1 for shape(n_task_dims, n_steps) x0, g = np.zeros(1), np.ones(1) beh = DMPBehavior(execution_time, dt, 20) beh.init(3, 3) beh.set_meta_parameters(["x0", "g"], [x0, g]) X_demo = np.empty((1, n_steps, 1)) X_demo[0, :, 0] = np.linspace(0, 1, n_steps) assert_equal(n_steps, X_demo.shape[1]) beh.imitate(X_demo, alpha=0.01) X, Xd, Xdd = beh.trajectory() assert_equal(X_demo[0, :].shape, X.shape)
def test_dmp_imitate_2d(): x0, g, execution_time, dt = np.zeros(2), np.ones(2), 1.0, 0.001 beh = DMPBehavior(execution_time, dt, 20) beh.init(6, 6) beh.set_meta_parameters(["x0", "g"], [x0, g]) X_demo = make_minimum_jerk(x0, g, execution_time, dt)[0] # Without regularization beh.imitate(X_demo) X = beh.trajectory()[0] assert_array_almost_equal(X_demo.T[0], X, decimal=2) # With alpha > 0 beh.imitate(X_demo, alpha=1.0) X = beh.trajectory()[0] assert_array_almost_equal(X_demo.T[0], X, decimal=3) # Self-imitation beh.imitate(X.T[:, :, np.newaxis]) X2 = beh.trajectory()[0] assert_array_almost_equal(X2, X, decimal=3)
label="Demostrated trajectory") plt.plot(np.linspace(0, 1, X_demo.shape[1]), 2 * X_demo[0, :], 'g--', linewidth=4, label="Demostrated trajectory") for ni in range(10, 50, 10): dmp = DMPBehavior( execution_time, dt, n_features=ni ) # Can be used to optimize the weights of a DMP with a black box optimizer. # Only the weights of the DMP will be optimized. We will use n_features gausssians. dmp.init(3, 3) #1*3 inputs and 1*3 outputs dmp.set_meta_parameters( ["x0", "g"], [x0, g ]) # Set the dmp metaparameters initial state x0 and goal state g. dmp.imitate(X_demo) # Learn weights of the DMP from a demonstration. plt.plot(np.linspace(0, 1, X_demo.shape[1]), dmp.trajectory()[0], '--', c=generateRandomColor(), label=str(ni) + "basis functions") plt.plot(mp_new_high[:, 0], mp_new_high[:, 1], '--g', label='Park et al.') plt.plot(mp_new_high[:, 0], mp_new_high[:, 2], '--g', label='Park et al.') plt.legend(loc="upper left") plt.show()
# Compute dmp in the original state space ni = 10 dmp = DMPBehavior(execution_time, dt, n_features=ni) # Can be used to optimize the weights of a DMP with a black box optimizer. # Only the weights of the DMP will be optimized. We will use n_features gausssians. dmp.init(6,6) #1*3 inputs and 1*3 outputs print("x0.shape", x0.shape) print("g.shape", g.shape) print("y_demo.shape", y_demo.shape) x0 = np.squeeze(x0) g = np.squeeze(g) dmp.set_meta_parameters(["x0", "g"], [x0, g]) # Set the dmp metaparameters initial state x0 and goal state g. y_demo = np.expand_dims(y_demo, axis=2) dmp.imitate(y_demo) # Learn weights of the DMP from a demonstration. trajectory_goal_demo = dmp.trajectory() # New goal g [0] = 10*np.pi/180 g [1] = -10*np.pi/180 dmp.set_meta_parameters(["x0", "g"], [x0, g]) # Set the dmp metaparameters initial state x0 and goal state g. dmp.imitate(y_demo) # Learn weights of the DMP from a demonstration. trajectory_new_goal_demo = dmp.trajectory() # Compute dmp in the transformation state space ni = 10 dmpNewSpace = DMPBehavior(execution_time, dt, n_features=ni) # Can be used to optimize the weights of a DMP with a black box optimizer. # Only the weights of the DMP will be optimized. We will use n_features gausssians.
def dmp_to_trajectory(dmp, x0, g, gd, execution_time): """Computes trajectory generated by open-loop controlled DMP.""" dmp.set_meta_parameters(["x0", "g", "gd", "execution_time"], [x0, g, gd, execution_time]) return dmp.trajectory() x0 = np.zeros(2) g = np.ones(2) dt = 0.001 execution_time = 1.0 dmp = DMPBehavior(execution_time, dt, n_features=20) dmp.init(6, 6) dmp.set_meta_parameters(["x0", "g"], [x0, g]) X_demo = make_minimum_jerk(x0, g, execution_time, dt)[0] dmp.imitate(X_demo) plt.figure() plt.subplots_adjust(wspace=0.3, hspace=0.6) for gx in np.linspace(0.5, 1.5, 6): g_new = np.array([gx, 1.0]) X, Xd, Xdd = dmp_to_trajectory(dmp, x0, g_new, np.zeros(2), 1.0) ax = plt.subplot(321) ax.set_title("Goal adaption") ax.set_xlabel("$x_1$") ax.set_ylabel("$x_2$") ax.plot(X[:, 0], X[:, 1]) ax = plt.subplot(322)