Пример #1
0
def world_loop(opts_dict):
    params = {
            'spawn': 15,
            'weather': 'clear_noon',
            'n_vehicles': 0
            }

    with cu.CarlaWrapper(TOWN, cu.VEHICLE_NAME, PORT) as env:
        env.init(**params)
        agent = RoamingAgent(env._player, False, opts_dict)

        # Hack: fill up controller experience.
        for _ in range(30):
            env.tick()
            env.apply_control(agent.run_step()[0])

        for _ in tqdm.tqdm(range(125)):
            env.tick()

            observations = env.get_observations()
            inputs = cu.get_inputs(observations)

            debug = dict()
            control, command = agent.run_step(inputs, debug_info=debug)
            env.apply_control(control)

            observations.update({'control': control, 'command': command})

            processed = cu.process(observations)

            yield debug

            bzu.show_image('rgb', processed['rgb'])
            bzu.show_image('birdview', cu.visualize_birdview(processed['birdview']))
Пример #2
0
def get_episode(env, params):
    data = list()
    progress = tqdm.tqdm(range(params.frames_per_episode), desc='Frame')
    start, target = env.pose_tasks[np.random.randint(len(env.pose_tasks))]
    env_params = {
        'weather': np.random.choice(list(cu.TRAIN_WEATHERS.keys())),
        'start': start,
        'target': target,
        'n_pedestrians': params.n_pedestrians,
        'n_vehicles': params.n_vehicles,
    }

    env.init(**env_params)
    env.success_dist = 5.0

    agent = NoisyAgent(env)
    agent.set_route(env._start_pose.location, env._target_pose.location)

    # Real loop.
    while len(data) < params.frames_per_episode and not env.is_success(
    ) and not env.collided:
        for _ in range(params.frame_skip):
            env.tick()

            observations = env.get_observations()
            control, command, last_status, real_control = agent.run_step(
                observations)
            agent_debug = agent.debug
            env.apply_control(control)

            observations['command'] = command
            observations['control'] = control
            observations['real_control'] = real_control

            if not params.nodisplay:
                _debug(observations, agent_debug)

        observations['control'] = real_control
        processed = cu.process(observations)

        data.append(processed)

        progress.update(1)

    progress.close()

    if (not env.is_success() and not env.collided) or len(data) < 500:
        return None

    return data
Пример #3
0
def _debug(observations, agent_debug):
    import cv2

    processed = cu.process(observations)

    control = observations['control']
    control = [control.steer, control.throttle, control.brake]
    control = ' '.join(str('%.2f' % x).rjust(5, ' ') for x in control)
    real_control = observations['real_control']
    real_control = [
        real_control.steer, real_control.throttle, real_control.brake
    ]
    real_control = ' '.join(
        str('%.2f' % x).rjust(5, ' ') for x in real_control)

    canvas = np.uint8(observations['rgb']).copy()
    rows = [x * (canvas.shape[0] // 10) for x in range(10 + 1)]
    cols = [x * (canvas.shape[1] // 10) for x in range(10 + 1)]

    WHITE = (255, 255, 255)
    CROP_SIZE = 192
    X = 176
    Y = 192 // 2
    R = 2

    def _write(text, i, j):
        cv2.putText(canvas, text, (cols[j], rows[i]), cv2.FONT_HERSHEY_SIMPLEX,
                    0.4, WHITE, 1)

    _command = {
        1: 'LEFT',
        2: 'RIGHT',
        3: 'STRAIGHT',
        4: 'FOLLOW',
    }.get(int(observations['command']), '???')

    _write('Command: ' + _command, 1, 0)
    _write('Velocity: %.1f' % np.linalg.norm(observations['velocity']), 2, 0)
    _write('Real: %s' % control, -5, 0)
    _write('Control: %s' % control, -4, 0)

    r = 2
    birdview = cu.visualize_birdview(crop_birdview(processed['birdview']))

    def _dot(x, y, color):
        x = int(x)
        y = int(y)
        birdview[176 - r - x:176 + r + 1 - x,
                 96 - r + y:96 + r + 1 + y] = color

    _dot(0, 0, [255, 255, 255])

    ox, oy = observations['orientation']
    R = np.array([[ox, oy], [-oy, ox]])

    u = np.array(agent_debug['waypoint']) - np.array(agent_debug['vehicle'])
    u = R.dot(u[:2])
    u = u * 4

    _dot(u[0], u[1], [255, 255, 255])

    def _stick_together(a, b):
        h = min(a.shape[0], b.shape[0])

        r1 = h / a.shape[0]
        r2 = h / b.shape[0]

        a = cv2.resize(a, (int(r1 * a.shape[1]), int(r1 * a.shape[0])))
        b = cv2.resize(b, (int(r2 * b.shape[1]), int(r2 * b.shape[0])))

        return np.concatenate([a, b], 1)

    full = _stick_together(canvas, birdview)

    bu.show_image('full', full)
def get_episode(env, params):
    data = list()
    progress = tqdm.tqdm(range(params.frames_per_episode), desc='Frame')
    start, target = env.pose_tasks[np.random.randint(len(env.pose_tasks))]
    env_params = {
        'weather': np.random.choice(list(cu.TRAIN_WEATHERS.keys())),
        'start': start,
        'target': target,
        'n_pedestrians': params.n_pedestrians,
        'n_vehicles': params.n_vehicles,
    }

    env.init(**env_params)
    env.success_dist = 5.0

    agent = NoisyAgent(env)
    agent.set_route(env._start_pose.location, env._target_pose.location)

    world = env._client.get_world()
    traffic_light_list = world.get_actors().filter('*traffic_light*')

    # Real loop.
    while len(data) < params.frames_per_episode and not env.is_success(
    ) and not env.collided:
        for _ in range(params.frame_skip):
            env.tick()

            observations = env.get_observations()
            control, command, last_status, real_control = agent.run_step(
                observations)
            agent_debug = agent.debug
            env.apply_control(control)

            observations['command'] = command
            observations['control'] = control
            observations['real_control'] = real_control

            # Traffic lights
            is_intersection_marin, traffic_light, light_state, distance_to_traffic_light = \
                agent.detect_traffic_light(traffic_light_list)
            if traffic_light:
                print("INTERSECTION MARIN = ", is_intersection_marin)
                print(
                    "incoming traffic light is at " +
                    str(distance_to_traffic_light) + "meters and is of color ",
                    agent.LightState(light_state).name)

                # visualise traffic lights with a green rectangle
                # green = carla.Color(0, 255, 0)
                # size_plot_point = 0.1
                # life_time_plot_point = 1.0
                # world.debug.draw_point(traffic_light.get_location(), color=green, size=size_plot_point * 10,
                #                        life_time=life_time_plot_point)

                observations['is_traffic_light'] = True
                observations['traffic_light_color'] = light_state - 1
                observations[
                    'traffic_light_distance'] = distance_to_traffic_light

            else:
                observations['is_traffic_light'] = False
                observations['traffic_light_color'] = 0
                observations['traffic_light_distance'] = 0
            print(f'is traffic light: {observations["is_traffic_light"]}')
            print(f'traffic color: {observations["traffic_light_color"]}')
            print(
                f'traffic distance: {observations["traffic_light_distance"]}')

            if not params.nodisplay:
                _debug(observations, agent_debug)

        observations['control'] = real_control
        processed = cu.process(observations)

        data.append(processed)

        progress.update(1)

    progress.close()

    if (not env.is_success() and not env.collided) or len(data) < 500:
        return None

    return data