Esempio n. 1
0
def generate_obstacle(rng,
                      path,
                      vessel,
                      displacement_dist_std=150,
                      obst_radius_distr=np.random.poisson,
                      obst_radius_mean=30):
    min_distance = 0
    while min_distance <= 0:
        obst_displacement_dist = np.random.normal(0, displacement_dist_std)
        obst_arclength = (0.1 + 0.8 * rng.rand()) * path.length
        obst_position = path(obst_arclength)
        obst_displacement_angle = geom.princip(
            path.get_direction(obst_arclength) - np.pi / 2)
        obst_position += obst_displacement_dist * np.array(
            [np.cos(obst_displacement_angle),
             np.sin(obst_displacement_angle)])
        obst_radius = max(1, obst_radius_distr(obst_radius_mean))

        vessel_distance_vec = geom.Rzyx(0, 0, -vessel.heading).dot(
            np.hstack([obst_position - vessel.position, 0]))
        vessel_distance = np.linalg.norm(
            vessel_distance_vec) - vessel.width - obst_radius
        goal_distance = np.linalg.norm(obst_position -
                                       path(path.length)) - obst_radius
        min_distance = min(vessel_distance, goal_distance)

    return (obst_position, obst_radius)
Esempio n. 2
0
    def update_control_errors(self):
        # Update cruise speed error
        self.u_error = np.clip(
            (self.cruise_speed - self.vessel.relative_velocity[0]) / 2, -1, 1)
        self.chi_error = 0.0
        self.e = 0.0
        self.upsilon_error = 0.0
        self.h = 0.0

        # Get path course and elevation
        s = self.prog
        chi_p, upsilon_p = self.path.get_direction_angles(s)

        # Calculate tracking errors
        SF_rotation = geom.Rzyx(0, upsilon_p, chi_p)
        epsilon = np.transpose(SF_rotation).dot(self.vessel.position -
                                                self.path(self.prog))
        e = epsilon[1]
        h = epsilon[2]

        # Calculate course and elevation errors from tracking errors
        chi_r = np.arctan2(-e, self.la_dist)
        upsilon_r = np.arctan2(h, np.sqrt(e**2 + self.la_dist**2))
        chi_d = chi_p + chi_r
        upsilon_d = upsilon_p + upsilon_r
        self.chi_error = np.clip(
            geom.ssa(self.vessel.chi - chi_d) / np.pi, -1, 1)
        #self.e = np.clip(e/12, -1, 1)
        self.e = e
        self.upsilon_error = np.clip(
            geom.ssa(self.vessel.upsilon - upsilon_d) / np.pi, -1, 1)
        #self.h = np.clip(h/12, -1, 1)
        self.h = h
Esempio n. 3
0
    def _sim(self):
        psi = self._state[2]
        nu = self._state[3:]

        eta_dot = geom.Rzyx(0, 0, geom.princip(psi)).dot(nu)
        nu_dot = const.M_inv.dot(
            const.B(nu).dot(self.input) - const.D(nu).dot(nu) -
            const.C(nu).dot(nu) - const.L(nu).dot(nu))
        state_dot = np.concatenate([eta_dot, nu_dot])
        self._state += state_dot * self.t_step
        self._state[2] = geom.princip(self._state[2])
Esempio n. 4
0
    def _state_dot(self, state):
        psi = state[2]
        nu = state[3:]

        tau = np.array([self.input[0], 0, self.input[1]])

        eta_dot = geom.Rzyx(0, 0, geom.princip(psi)).dot(nu)
        nu_dot = const.M_inv.dot(tau
                                 #- const.D.dot(nu)
                                 - const.N(nu).dot(nu))
        state_dot = np.concatenate([eta_dot, nu_dot])
        return state_dot
Esempio n. 5
0
    def observe(self):
        """
        Generates the observation of the environment.
        Parameters
        ----------
        action : np.array
            [propeller_input, rudder_position].

        Returns
        -------
        obs : np.array
            [
            surge velocity,
            sway velocity,
            heading error,
            distance to goal,
            propeller_input,
            rudder_positione,
            self.nsectors*[closeness to closest obstacle in sector]
            ]
            All observations are between -1 and 1.
        """
        obst_range = self.config["obst_detection_range"]

        goal_vector = self.goal - self.vessel.position
        goal_direction = np.arctan2(goal_vector[1], goal_vector[0])
        heading_error = float(
            geom.princip(goal_direction - self.vessel.heading))
        obs = np.zeros((self.nstates + self.nsectors, ))

        obs[0] = np.clip(self.vessel.velocity[0] / self.vessel.max_speed, -1,
                         1)
        obs[1] = np.clip(self.vessel.velocity[1] / 0.26, -1, 1)
        obs[2] = np.clip(self.vessel.yawrate / 0.55, -1, 1)
        obs[3] = np.clip(heading_error / np.pi, -1, 1)
        for obst in self.obstacles:
            distance_vec = geom.Rzyx(0, 0, -self.vessel.heading).dot(
                np.hstack([obst.position - self.vessel.position, 0]))
            dist = linalg.norm(distance_vec)
            if dist < obst_range + obst.radius + self.vessel.radius:
                ang = ((float(np.arctan2(distance_vec[1], distance_vec[0])) +
                        np.pi) / (2 * np.pi))
                closeness = 1 - np.clip(
                    (dist - self.vessel.radius - obst.radius) / obst_range, 0,
                    1)
                isector = (self.nstates + int(np.floor(ang * self.nsectors)))
                if isector == self.nstates + self.nsectors:
                    isector = self.nstates
                if obs[isector] < closeness:
                    obs[isector] = closeness
        return obs
Esempio n. 6
0
    def observe(self):
        """
        Generates the observation of the environment.
        Parameters
        ----------
        action : np.array
            [propeller_input, rudder_position].

        Returns
        -------
        obs : np.array
            [
            surge velocity,
            sway velocity,
            yawrate,
            heading error,
            cross track error,
            propeller_input,
            rudder_position,
            ]
            All observations are between -1 and 1.
        """
        la_heading = self.path.get_direction(
            self.path_prog[-1] + self.config["la_dist"])
        heading_error_la = float(geom.princip(la_heading
                                              - self.vessel.heading))
        path_position = (self.path(self.path_prog[-1]
                                   + self.config["la_dist"])
                         - self.vessel.position)
        target_heading = np.arctan2(path_position[1], path_position[0])
        heading_error = float(geom.princip(target_heading
                                           - self.vessel.heading))
        path_direction = self.path.get_direction(self.path_prog[-1])
        cross_track_error = geom.Rzyx(0, 0, -path_direction).dot(
            np.hstack([self.path(self.path_prog[-1])
                       - self.vessel.position, 0]))[1]

        obs = np.zeros((self.nstates,))

        obs[0] = np.clip(self.vessel.velocity[0]
                         /self.vessel.max_speed, -1, 1)
        obs[1] = np.clip(self.vessel.velocity[1]
                         /0.26, -1, 1)
        obs[2] = np.clip(self.vessel.yawrate/0.55, -1, 1)
        obs[3] = np.clip(heading_error_la/np.pi, -1, 1)
        obs[4] = np.clip(heading_error/np.pi, -1, 1)
        obs[5] = np.clip(cross_track_error
                         /25, -1, 1)

        return obs
Esempio n. 7
0
    def __call__(self, state):
        """Returns the current velcotiy in {b} to use in in AUV kinematics"""
        phi = state[3]
        theta = state[4]
        psi = state[5]
        omega_body = state[9:12]

        vel_current_NED = np.array([self.Vc*np.cos(self.alpha)*np.cos(self.beta), self.Vc*np.sin(self.beta), self.Vc*np.sin(self.alpha)*np.cos(self.beta)])
        vel_current_BODY = np.transpose(geom.Rzyx(phi, theta, psi)).dot(vel_current_NED)
        vel_current_BODY_dot = -geom.S_skew(omega_body).dot(vel_current_BODY)

        nu_c = np.array([*vel_current_BODY, 0, 0, 0])
        nu_c_dot = np.array([*vel_current_BODY_dot, 0, 0, 0])

        return nu_c
Esempio n. 8
0
 def update_nearby_obstacles(self):
     """
     Updates the nearby_obstacles array.
     """
     self.nearby_obstacles = []
     for obstacle in self.obstacles:
         distance_vec_NED = obstacle.position - self.vessel.position
         distance = np.linalg.norm(distance_vec_NED)
         distance_vec_BODY = np.transpose(
             geom.Rzyx(*self.vessel.attitude)).dot(distance_vec_NED)
         heading_angle_BODY = np.arctan2(distance_vec_BODY[1],
                                         distance_vec_BODY[0])
         pitch_angle_BODY = np.arctan2(
             distance_vec_BODY[2],
             np.sqrt(distance_vec_BODY[0]**2 + distance_vec_BODY[1]**2))
         # check if the obstacle is inside the sonar window
         if distance - self.vessel.safety_radius - obstacle.radius <= self.sonar_range and abs(heading_angle_BODY) <= self.sensor_span[0]*np.pi/180 \
         and abs(pitch_angle_BODY) <= self.sensor_span[1]*np.pi/180:
             self.nearby_obstacles.append(obstacle)
         elif distance <= obstacle.radius + self.vessel.safety_radius:
             self.nearby_obstacles.append(obstacle)
Esempio n. 9
0
    def navigate(self, path: Path) -> np.ndarray:
        """
        Calculates and returns navigation states representing the vessel's attitude
        with respect to the desired path.

        Returns
        -------
        navigation_states : np.ndarray
        """

        # Calculating path arclength at reference point, i.e. the point closest to the vessel
        vessel_arclength = path.get_closest_arclength(self.position)

        # Calculating tangential path direction at reference point
        path_direction = path.get_direction(vessel_arclength)
        cross_track_error = geom.Rzyx(0, 0, -path_direction).dot(
            np.hstack([path(vessel_arclength) - self.position, 0]))[1]

        # Calculating tangential path direction at look-ahead point
        target_arclength = min(
            path.length, vessel_arclength + self.config["look_ahead_distance"])
        look_ahead_path_direction = path.get_direction(target_arclength)
        look_ahead_heading_error = float(
            geom.princip(look_ahead_path_direction - self.heading))

        # Calculating vector difference between look-ahead point and vessel position
        target_vector = path(target_arclength) - self.position

        # Calculating heading error
        target_heading = np.arctan2(target_vector[1], target_vector[0])
        heading_error = float(geom.princip(target_heading - self.heading))

        # Calculating path progress
        progress = vessel_arclength / path.length
        self._progress = progress

        # Concatenating states
        self._last_navi_state_dict = {
            'surge_velocity': self.velocity[0],
            'sway_velocity': self.velocity[1],
            'yaw_rate': self.yaw_rate,
            'look_ahead_heading_error': look_ahead_heading_error,
            'heading_error': heading_error,
            'cross_track_error': cross_track_error / 100,
            'target_heading': target_heading,
            'look_ahead_path_direction': look_ahead_path_direction,
            'path_direction': path_direction,
            'vessel_arclength': vessel_arclength,
            'target_arclength': target_arclength
        }
        navigation_states = np.array([
            self._last_navi_state_dict[state]
            for state in Vessel.NAVIGATION_FEATURES
        ])

        # Deciding if vessel has reached the goal
        goal_distance = linalg.norm(path.end - self.position)

        reached_goal = goal_distance <= self.config[
            "min_goal_distance"] or progress >= self.config["min_path_progress"]
        self._reached_goal = reached_goal

        return navigation_states
Esempio n. 10
0
    def observe(self):
        """
        Generates the observation of the environment.
        Parameters
        ----------
        action : np.array
            [propeller_input, rudder_position].

        Returns
        -------
        obs : np.array
            [
            surge velocity,
            sway velocity,
            yawrate,
            heading error,
            cross track error,
            propeller_input,
            rudder_position,
            ]
            All observations are between -1 and 1.
        """
        la_heading = self.path.get_direction(self.path_prog[-1] +
                                             self.config["la_dist"])
        heading_error_la = float(geom.princip(la_heading -
                                              self.vessel.heading))
        path_position = (
            self.path(self.path_prog[-1] + self.config["la_dist"]) -
            self.vessel.position)
        target_heading = np.arctan2(path_position[1], path_position[0])
        heading_error = float(
            geom.princip(target_heading - self.vessel.heading))
        path_direction = self.path.get_direction(self.path_prog[-1])
        cross_track_error = geom.Rzyx(0, 0, -path_direction).dot(
            np.hstack(
                [self.path(self.path_prog[-1]) - self.vessel.position, 0]))[1]

        obs = np.zeros((self.nstates + self.nsectors, ))

        obs[0] = np.clip(self.vessel.velocity[0] / self.vessel.max_speed, -1,
                         1)
        obs[1] = np.clip(self.vessel.velocity[1] / 0.26, -1, 1)
        obs[2] = np.clip(self.vessel.yawrate / 0.55, -1, 1)
        obs[3] = np.clip(heading_error_la / np.pi, -1, 1)
        obs[4] = np.clip(heading_error / np.pi, -1, 1)
        obs[5] = np.clip(cross_track_error / 50, -10000, 10000)

        obst_range = self.config["obst_detection_range"]
        for obst in self.obstacles:
            distance_vec = geom.Rzyx(0, 0, -self.vessel.heading).dot(
                np.hstack([obst.position - self.vessel.position, 0]))
            dist = linalg.norm(distance_vec)
            if dist < obst_range + obst.radius + self.vessel.radius:
                ang = ((float(np.arctan2(distance_vec[1], distance_vec[0])) +
                        np.pi) / (2 * np.pi))
                closeness = 1 - np.clip(
                    (dist - self.vessel.radius - obst.radius) / obst_range, 0,
                    1)
                isector = (self.nstates + int(np.floor(ang * self.nsectors)))
                if isector == self.nstates + self.nsectors:
                    isector = self.nstates
                if obs[isector] < closeness:
                    obs[isector] = closeness

        return obs