def integrate(p0, R0, pdot0, om0, t_out): ''' integrate Newton-Euler equations, then integrate kinematics t_out = times at which to return states ''' # setup t0 = t_out[0] tf_out = t_out[-1] n_t_out = t_out.size # integration times dt = .001 t = np.arange(t0, tf_out + dt, dt) # tf_out+dt: make sure tf_out is included tf = t[-1] n_t = t.size # integrate newton-euler (state is pdot & om) pdot_om0 = np.concatenate((pdot0, om0)) sol = sp.integrate.solve_ivp(newton_euler, (t0, tf), pdot_om0, method='RK45', t_eval=t) pdot_om = sol.y pdot = pdot_om[:3, :] om = pdot_om[3:, :] # 1st-order forward Euler integration of kinematics to get p and R p = np.full((3, n_t), np.nan) R = np.full((3, 3, n_t), np.nan) p[:, 0] = p0 R[:, :, 0] = R0 for i in range(1, n_t): p[:, i] = p[:, i - 1] + dt * pdot[:, i - 1] R[:, :, i] = R[:, :, i - 1] @ so3.exp(so3.cross(dt * om[:, i - 1])) # get tangent space representation of R so we can interpolate s = np.full((3, n_t), np.nan) for i in range(n_t): s[:, i] = so3.skew_elements(so3.log(R[:, :, i])) # interpolate t values using state x=(p,s,v,om) x = np.concatenate((p, s, pdot, om), axis=0) interp_fun = sp.interpolate.interp1d(t, x) x_out = interp_fun(t_out) # get x values p_out = x_out[:3, :] s_out = x_out[3:6, :] pdot_out = x_out[6:9, :] om_out = x_out[9:, :] # convert tangent space s back into rotations R_out = np.full((3, 3, n_t_out), np.nan) for i in range(n_t_out): R_out[:, :, i] = so3.exp(so3.cross(s_out[:, i])) return p_out, R_out, pdot_out, om_out
def gen_and_eval(mode): # to be filled p = np.full((3, n_ims), np.nan) R = np.full((3, 3, n_ims), np.nan) if mode == 'trans_x': x = np.linspace(-.2, .2, n_ims) for i in range(n_ims): R[:, :, i] = R_upright p[:, i] = np.array([x[i], .8, 0]) img_dir = dirs.trans_x_dir elif mode == 'trans_y': y = np.linspace(.4, 1.8, n_ims) for i in range(n_ims): R[:, :, i] = R_upright p[:, i] = np.array([0, y[i], 0]) img_dir = dirs.trans_y_dir elif mode == 'trans_z': z = np.linspace(-.15, .15, n_ims) for i in range(n_ims): R[:, :, i] = R_upright p[:, i] = np.array([0, .8, z[i]]) img_dir = dirs.trans_z_dir elif mode == 'rot_x': # rotate eps is needed to avoided wrapping of the exp/log maps ang = np.linspace(-np.pi + rot_eps, np.pi - rot_eps, n_ims) for i in range(n_ims): s_i = np.array([ang[i], 0, 0]) R[:, :, i] = so3.exp(so3.cross(s_i)) @ R_upright p[:, i] = np.array([0, .8, 0]) img_dir = dirs.rot_x_dir elif mode == 'rot_y': ang = np.linspace(-np.pi + rot_eps, np.pi - rot_eps, n_ims) for i in range(n_ims): s_i = np.array([0, ang[i], 0]) R[:, :, i] = so3.exp(so3.cross(s_i)) @ R_upright p[:, i] = np.array([0, .8, 0]) img_dir = dirs.rot_y_dir elif mode == 'rot_z': ang = np.linspace(-np.pi + rot_eps, np.pi - rot_eps, n_ims) for i in range(n_ims): s_i = np.array([0, 0, ang[i]]) R[:, :, i] = so3.exp(so3.cross(s_i)) @ R_upright p[:, i] = np.array([0, .8, 0]) img_dir = dirs.rot_z_dir # render and predict dt = 1 # this doesn't matter for this simulation br.soup_gen(dt, p, R, img_dir, lighting_energy, world_RGB) p, R, p_est, R_est = db.get_predictions(img_dir, print_errors=False)
def h(p_prev, R_prev, u, v): ''' measurement function ''' p_new = p_prev + v[:3] R_noise = so3.exp(so3.cross(v[3:])) R_new = R_prev @ R_noise return p_new, R_new
def filter(p0_hat, R0_hat, pdot0_hat, om0_hat, COV_xx_0_hat, COV_ww, COV_vv, U, P_MEAS, R_MEAS, dt): ''' Unscented Filter from Section 3.7 of Optimal Estimation of Dynamic Systems (2nd ed.) by Crassidis & Junkins ''' # system constants alpha = 1e-2 beta = 2 kappa = 2 n = 12 n_w = COV_ww.shape[0] n_v = COV_vv.shape[0] n_t = P_MEAS.shape[1] # number of time points n_y = 6 # constants L = n + n_w + n_v # size of augmented state lam = (alpha**2) * (L + kappa) - L # eq. 3.256 gam = np.sqrt(L + lam) # eq. 3.255 # weights W_mean = np.full(2 * L + 1, np.nan) W_cov = np.full(2 * L + 1, np.nan) W_mean[0] = lam / (L + lam) W_cov[0] = W_mean[0] + (1 - alpha**2 + beta) W_mean[1:] = 1 / (2 * (L + lam)) W_cov[1:] = 1 / (2 * (L + lam)) # create arrays for saving values in the for loop P_HAT = np.full((3, n_t), np.nan) R_HAT = np.full((3, 3, n_t), np.nan) PDOT_HAT = np.full((3, n_t), np.nan) OM_HAT = np.full((3, n_t), np.nan) COV_XX_ALL = np.full((n, n, n_t), np.nan) Y_HAT = np.full((n_y, 2 * L + 1), np.nan) # y_hat^(i) for i=0,...,2L # fill arraysw with initial estimates P_HAT[:, 0] = p0_hat R_HAT[:, :, 0] = R0_hat PDOT_HAT[:, 0] = pdot0_hat OM_HAT[:, 0] = om0_hat s0_hat = np.array([0, 0, 0]) # initialize variables to be used in for loop x_hat = np.block([p0_hat, s0_hat, pdot0_hat, om0_hat]) x_hat = x_hat[:, np.newaxis] COV_xx = COV_xx_0_hat R_hat = R0_hat # iterate over all time points for i in range(1, n_t): # load control and measurement at time i u = U[:, i] s_meas = so3.skew_elements(so3.log(R_hat.T @ R_MEAS[:, :, i])) # eq. 5 y = np.block([P_MEAS[:, i], s_meas]) y = y[:, np.newaxis] # cross-correlations (usually zero, see Crassidis after eq. 3.252) COV_xw = np.full((n, n_w), 0.0) COV_xv = np.full((n, n_v), 0.0) COV_wv = np.full((n_w, n_v), 0.0) # sigma points COV_chi = np.block([[COV_xx, COV_xw, COV_xv], [COV_xw.T, COV_ww, COV_wv], [COV_xv.T, COV_wv.T, COV_vv]]) M = gam * np.linalg.cholesky( COV_chi) # returns L such that COV_chi = L @ L.T sig = np.block([M, -M]) chi_hat_pls = np.block([[x_hat], [np.zeros((n_w + n_v, 1))]]) CHI_HAT = np.block([chi_hat_pls, chi_hat_pls + sig ]) # all chi_hat^(i) for i=0,...,2L X_HAT = CHI_HAT[:n, :] # all x_hat^(i) for i=0,...,2L W_HAT = CHI_HAT[n:n + n_w, :] # all w_hat^(i) for i=0,...,2L V_HAT = CHI_HAT[n + n_w:, :] # all v_hat^(i) for i=0,...,2L # iteration (over all sigma points) for k in range(2 * L + 1): # dynamics om_k = X_HAT[9:, k] R_k = R_hat @ so3.exp(so3.cross(X_HAT[3:6, k])) p_new, R_new, v_new, om_new = f(X_HAT[:3, k], R_k, X_HAT[6:9, k], om_k, W_HAT[:, k], u, dt) s_new = so3.skew_elements(so3.log(R_hat.T @ R_new)) X_HAT[:, k] = np.block([p_new, s_new, v_new, om_new]) # measurement p_meas, R_meas = h(p_new, R_new, u, V_HAT[:, k]) s_meas = so3.skew_elements(so3.log(R_hat.T @ R_meas)) Y_HAT[:, k] = np.block([p_meas, s_meas]) # predictions x_hat = np.sum(W_mean * X_HAT, axis=1) x_hat = x_hat[:, np.newaxis] COV_xx = W_cov * (X_HAT - x_hat) @ (X_HAT - x_hat).T # COV_xx minus y_hat = np.sum(W_mean * Y_HAT, axis=1) # eq. 2.262 y_hat = y_hat[:, np.newaxis] # covariances COV_yy = W_cov * (Y_HAT - y_hat) @ (Y_HAT - y_hat).T COV_xy = W_cov * (X_HAT - x_hat) @ (Y_HAT - y_hat).T # update e = y - y_hat # eq. 3.250 K = COV_xy @ np.linalg.inv(COV_yy) # eq. 3.251 x_hat = x_hat + K @ e # eq. 3.249a COV_xx = COV_xx - K @ COV_yy @ K.T # eq. 3.249b # rotation R_hat = R_hat @ so3.exp(so3.cross(x_hat[3:6].ravel())) x_hat[3:6] = 0 # save values P_HAT[:, i] = x_hat[:3].ravel() R_HAT[:, :, i] = R_hat PDOT_HAT[:, i] = x_hat[6:9].ravel() OM_HAT[:, i] = x_hat[9:].ravel() COV_XX_ALL[:, :, i] = COV_xx return P_HAT, R_HAT, PDOT_HAT, OM_HAT, COV_XX_ALL
R0 = t3d.quaternions.quat2mat(q0) pdot0 = np.array([0.9, 1.1, 2.3]) om0 = np.array([5, 8, 4]) # rigid body dynamics p, R, pdot, om = rb.integrate(p0, R0, pdot0, om0, t) save_dir = os.path.join(dirs.results_dir, 'filter_test/') # simulate measurements noise_p = .04 noise_R = .17 p_meas = p + noise_p*np.random.uniform(-1,1,(3,n_ims)) R_meas = np.full((3,3,n_ims), np.nan) for i in range(n_ims): s_noise = noise_R*np.random.uniform(.5,-.5,3) R_noise = so3.exp(so3.cross(s_noise)) R_meas[:,:,i] = R_noise @ R[:,:,i] # filter initial estimates p0_hat = p_meas[:,0] # use the true value to make it a fair comparison R0_hat = R_meas[:,:,0] # use the true value to make it a fair comparison pdot0_hat = pdot0 + .2*np.array([-1.1, 1.2, 1.1]) om0_hat = om0 + .6*np.array([-1.0, 1.1, 1.0]) cov_xx_0_hat = np.ones(12) COV_xx_0_hat = np.diag(cov_xx_0_hat) # run filter U = np.full((1, n_ims), 0.0) # there is no control in this problem COV_ww, COV_vv = df.get_noise_covariances() p_filt, R_filt, pdot_filt, om_filt, COV_XX_ALL = uf.filter(p0_hat, R0_hat, pdot0_hat, om0_hat, COV_xx_0_hat, COV_ww, COV_vv, U, p_meas, R_meas, dt)