We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Setup Describe the setup you are using to run CARLA along with its version:
Describe the bug I am using the following env.py from env import routes
import random import re import math import os import queue import numpy as np import matplotlib.pyplot as plt import time import carla from carla_agents.navigation.basic_agent import BasicAgent from carla_agents.navigation.local_planner import RoadOption from carla_agents.navigation.global_route_planner import GlobalRoutePlanner class CarlaEnv: def __init__(self, world='Town02', fps=10, image_w=256, image_h=112, evaluate=False, on_test_set=False, eval_image_w=720, eval_image_h=480): self.image_w = image_w self.image_h = image_h self.evaluate = evaluate self.eval_image_w = eval_image_w self.eval_image_h = eval_image_h # episode variables self.max_episode_steps = 150 if not evaluate else 250 self.episode_number = -2 self.obs_number = 0 self.current_route = 0 # spawn and destination location self.on_test_set = on_test_set if not on_test_set: self.spawns = routes.train_spawns self.rotations = routes.train_rotations self.dests = routes.train_dests else: self.spawns = routes.test_spawns self.rotations = routes.test_rotations self.dests = routes.test_dests # input/output dimension (3 RGB cameras and 3 actions) self.observation_space = (9, image_h, image_w) self.action_space = (3,) # connecting to carla self.client = carla.Client('localhost', 2000) self.client.set_timeout(5.0) # setting the world up self.world = self.client.load_world(world) print('carla connected.') self.world.tick() # set weather rgx = re.compile('.+?(?:(?<=[a-z])(?=[A-Z])|(?<=[A-Z])(?=[A-Z][a-z])|$)') name = lambda x: ' '.join(m.group(0) for m in rgx.finditer(x)) presets = [x for x in dir(carla.WeatherParameters) if re.match('[A-Z].+', x)] presets = [(getattr(carla.WeatherParameters, x), name(x)) for x in presets] self.world.set_weather(presets[6][0]) # fps settings settings = self.world.get_settings() settings.synchronous_mode = True settings.fixed_delta_seconds = 1 / fps self.world.apply_settings(settings) print("Reached line 68 in env.py") # calculate the route to destination (to calculate distance) world_map = self.world.get_map() planner = GlobalRoutePlanner(world_map, sampling_resolution=1.0) print("Reached line 72 in env.py") route = [] for i in range(len(self.spawns)): print("Reached line 75 in env.py") start_waypoint = world_map.get_waypoint(self.spawns[i]) end_waypoint = world_map.get_waypoint(self.dests[i]) route.append(planner.trace_route( start_waypoint.transform.location, end_waypoint.transform.location)) # calculate distance to destination for every point along the route self.distances = [[] for _ in range(len(self.spawns))] for i in range(len(self.spawns)): for j in range(len(route[i]) - 1): p1 = route[i][j][0].transform.location p2 = route[i][j+1][0].transform.location distance = math.sqrt((p1.x - p2.x) ** 2 + (p1.y - p2.y) ** 2) self.distances[i].append(distance) self.distances[i].append(0) for j in reversed(range(len(route[i]) - 1)): self.distances[i][j] += self.distances[i][j+1] # recording directory if self.evaluate: main_dir = 'agent_rollout' if not os.path.exists(main_dir): os.makedirs(main_dir) self.dir = os.path.join(main_dir, 'train' if not self.on_test_set else 'test') # sensor variables self.image_queues = [queue.Queue(), queue.Queue(), queue.Queue()] self.eval_image_queue = queue.Queue() self.vehicle = None self.agent = None self.cameras = [] self.eval_camera = None self.collision_sensor = None self.lane_invasion_sensor = None self.early_terminate = False self.invaded = False print("Reached line 113 in env.py") def reset(self, route=None): print("Reached line 116 in env.py") self.episode_number += 1 self.obs_number = 0 self.early_terminate = False self.invaded = False # choose a route based on episode number or the route passed as an argument if route is None: self.current_route = self.episode_number % len(self.spawns) else: self.current_route = route print("Reached line 127 in env.py") # deleting vehicle and sensors (if they already exist) self.image_queues = [queue.Queue(), queue.Queue(), queue.Queue()] self.destroy_sensors() self.cameras = [] # spawning vehicle blueprint_lib = self.world.get_blueprint_library() vehicle_bp = blueprint_lib.filter('model3')[0] vehicle_spawn_point = carla.Transform(self.spawns[self.current_route], self.rotations[self.current_route]) self.vehicle = self.world.spawn_actor(vehicle_bp, vehicle_spawn_point) # collision sensor self.collision_sensor = self.world.spawn_actor( blueprint_lib.find('sensor.other.collision'), carla.Transform(), attach_to=self.vehicle) self.collision_sensor.listen(lambda event: self.terminate()) print("Collision sensor spawned. Ticking world...") self.world.tick() print("Tick after collision sensor OK.") print("Reached line 144 in env.py") # lane invasion sensor self.lane_invasion_sensor = self.world.spawn_actor( blueprint_lib.find('sensor.other.lane_invasion'), carla.Transform(), attach_to=self.vehicle) self.lane_invasion_sensor.listen(lambda event: self.invade()) print("Lane invasion sensor spawned. Ticking world...") self.world.tick() print("Tick after lane invasion sensor OK.") # setting up main cameras bound_x = self.vehicle.bounding_box.extent.x bound_y = self.vehicle.bounding_box.extent.y bound_z = self.vehicle.bounding_box.extent.z print("Reached line 155 in env.py") # left camera camera_bp = blueprint_lib.find('sensor.camera.rgb') camera_bp.set_attribute('image_size_x', f'{self.image_w}') camera_bp.set_attribute('image_size_y', f'{self.image_h}') camera_bp.set_attribute('fov', str(120)) camera_spawn_point = carla.Transform( carla.Location(x=-0.2 * bound_x, y=-bound_y, z=0.4 * (0.5 + bound_z)), carla.Rotation(yaw=295, pitch=-20)) self.cameras.append(self.world.spawn_actor(camera_bp, camera_spawn_point, attach_to=self.vehicle, attachment_type=carla.AttachmentType.Rigid)) self.cameras[0].listen(self.image_queues[0].put) print("LEFT camera spawned. Ticking world...") self.world.tick() # <<<--- DOES IT CRASH HERE? print("Tick after LEFT camera OK.") # front camera camera_bp = blueprint_lib.find('sensor.camera.rgb') camera_bp.set_attribute('image_size_x', f'{self.image_w}') camera_bp.set_attribute('image_size_y', f'{self.image_h}') camera_spawn_point = carla.Transform( carla.Location(x=0.5 + bound_x, y=0.0, z=0.5 + bound_z), carla.Rotation(pitch=-15)) self.cameras.append(self.world.spawn_actor(camera_bp, camera_spawn_point, attach_to=self.vehicle, attachment_type=carla.AttachmentType.Rigid)) self.cameras[1].listen(self.image_queues[1].put) print("FRONT camera spawned. Ticking world...") self.world.tick() print("Tick after FRONT camera OK.") # right camera camera_bp = blueprint_lib.find('sensor.camera.rgb') camera_bp.set_attribute('image_size_x', f'{self.image_w}') camera_bp.set_attribute('image_size_y', f'{self.image_h}') camera_bp.set_attribute('fov', str(120)) camera_spawn_point = carla.Transform( carla.Location(x=-0.2 * bound_x, y=bound_y, z=0.4 * (0.5 + bound_z)), carla.Rotation(yaw=65, pitch=-20)) right_camera = self.world.spawn_actor(camera_bp, camera_spawn_point, attach_to=self.vehicle, attachment_type=carla.AttachmentType.Rigid) self.cameras.append(right_camera) self.cameras[2].listen(self.image_queues[2].put) print("RIGHT camera spawned. Ticking world...") self.world.tick() print("Tick after RIGHT camera OK.") print("Reached line 191 in env.py") # setting up evaluation camera if self.evaluate: camera_bp = blueprint_lib.find('sensor.camera.rgb') camera_bp.set_attribute('image_size_x', f'{self.eval_image_w}') camera_bp.set_attribute('image_size_y', f'{self.eval_image_h}') camera_spawn_point = carla.Transform( carla.Location(x=-2.0*(0.5+bound_x), y=0.0, z=2.0*(0.5+bound_z)), carla.Rotation(pitch=8.0) ) self.eval_camera = self.world.spawn_actor( camera_bp, camera_spawn_point, attach_to=self.vehicle, attachment_type=carla.AttachmentType.SpringArm ) self.eval_camera.listen(self.eval_image_queue.put) # episode record directory if self.evaluate and self.episode_number != -1: sub_dir = os.path.join(self.dir, f'ep_{self.episode_number}') if not os.path.exists(sub_dir): os.makedirs(sub_dir) print("Reached line 216 in env.py") # generate and process the first observation self.world.tick() # self.world.tick() # --- Simple loop: Tick until all queues have data --- print("Waiting for initial sensor data (simple loop)...") # Continue looping as long as ANY queue is empty while self.image_queues[0].empty() or \ self.image_queues[1].empty() or \ self.image_queues[2].empty(): # Tick the world to allow sensors to produce data print(" Ticking world, waiting for queues...") self.world.tick() # Optional tiny sleep to prevent potential high CPU if tick is extremely fast # import time time.sleep(0.001) print("All sensor queues have data now.") print("Reached line 219 in env.py") img_0 = self.process_image(self.image_queues[0].get(timeout=2)) img_1 = self.process_image(self.image_queues[1].get(timeout=2)) img_2 = self.process_image(self.image_queues[2].get(timeout=2)) obs = np.vstack((img_0, img_1, img_2)) print("Reached line 224 in env.py") if self.evaluate and self.episode_number != -1: image = self.process_image(self.eval_image_queue.get()) self.save_image(image) # setup agent to provide high-level commands self.agent = BasicAgent(self.vehicle) self.agent.ignore_traffic_lights(active=False) self.agent.set_destination(self.dests[self.current_route]) # initial command: move forward command = np.array([0.0, 1.0, 0.0]) # vehicle speed speed = 0 print("Reached line 235 in env.py") return obs, command, speed def step(self, action): # add noise to steer (used for testing robustness of agent) # action[1] = 0 # if random.random() < 0.25: # action[1] += 0.2 # clipping values # print("Reached line 246 in env.py") throttle = float(np.clip(action[0], 0, 1)) steer = float(np.clip(action[1], -1, 1)) brake = float(np.clip(action[2], 0, 1)) if brake < 0.01: brake = 0 # if self.obs_number < 50: # throttle, steer, brake = 0, 0, 0 # applying the action self.vehicle.apply_control( carla.VehicleControl(throttle=throttle, steer=steer, brake=brake)) self.world.tick() # get the next observation img_0 = self.process_image(self.image_queues[0].get(timeout=2.0)) img_1 = self.process_image(self.image_queues[1].get(timeout=2.0)) img_2 = self.process_image(self.image_queues[2].get(timeout=2.0)) obs = np.vstack((img_0, img_1, img_2)) if self.evaluate: image = self.process_image(self.eval_image_queue.get()) self.save_image(image) self.obs_number += 1 # get high-level command from agent's global planner _, road_option, num_points_done = self.agent.run_step() if road_option == RoadOption.LANEFOLLOW or road_option == RoadOption.STRAIGHT: command = np.array([0.0, 1.0, 0.0]) elif road_option == RoadOption.LEFT: command = np.array([1.0, 0.0, 0.0]) elif road_option == RoadOption.RIGHT: command = np.array([0.0, 0.0, 1.0]) # calculate distance to destination location = self.vehicle.get_transform().location aerial_dist = math.sqrt( (self.dests[self.current_route].x - location.x) ** 2 + (self.dests[self.current_route].y - location.y) ** 2) road_dist = self.distances[self.current_route][num_points_done] # check episode termination if self.obs_number == self.max_episode_steps or self.early_terminate: done = True info = {'distance': road_dist} elif aerial_dist < 5 or road_dist < 5: done = True info = {'distance': 0} else: done = False info = {} # environment reward (in case of collision or lane invasion, -25) if self.invaded or self.early_terminate: reward = -5 self.invaded = False else: reward = 0 # vehicle speed v = self.vehicle.get_velocity() speed = math.sqrt(v.x ** 2 + v.y ** 2 + v.z ** 2) # print(self.obs_number, command, road_dist) print("Reached line 312 in env.py") return obs, command, speed, reward, done, info # def destroy_sensors(self): # sensors = [*self.cameras, self.eval_camera, self.collision_sensor, self.lane_invasion_sensor] # for sensor in sensors: # if sensor is not None: # print(sensor) # sensor.stop() # sensor.destroy() # if self.vehicle: # self.vehicle.destroy() # self.world.tick() def destroy_sensors(self): print("Destroying sensors and vehicle...") # Use a client instance for batch destruction if preferred, or keep individual client = self.client # Get client reference if needed actors_to_destroy = [] sensors_to_stop = [] # Collect sensors that need stopping/destroying all_sensors = [*self.cameras, self.eval_camera, self.collision_sensor, self.lane_invasion_sensor] for sensor in all_sensors: if sensor is not None: # Check if alive before adding to list (optional robustness) try: if sensor.is_alive: sensors_to_stop.append(sensor) actors_to_destroy.append(sensor.id) else: # Already destroyed or invalid? print(f"Warning: Sensor {getattr(sensor, 'id', 'unknown')} found but not alive during destroy.") except Exception as e: print(f"Warning: Error checking sensor state during destroy: {e}") # Stop listeners first (important!) print(f"Stopping {len(sensors_to_stop)} sensors...") for sensor in sensors_to_stop: try: if sensor.is_listening: sensor.stop() except Exception as e: print(f"Warning: Error stopping sensor {sensor.id}: {e}") # Collect vehicle if alive if self.vehicle is not None: try: if self.vehicle.is_alive: actors_to_destroy.append(self.vehicle.id) else: print(f"Warning: Vehicle {self.vehicle.id} found but not alive during destroy.") except Exception as e: print(f"Warning: Error checking vehicle state during destroy: {e}") # Destroy actors using batch command if actors_to_destroy: print(f"Sending destroy commands for actors: {actors_to_destroy}") client.apply_batch_sync([carla.command.DestroyActor(actor_id) for actor_id in actors_to_destroy], True) print("Batch destroy command sent and acknowledged.") else: print("No live actors needed destruction.") # --- CRITICAL: Tick the world to process destruction --- print("Ticking world once to process destruction...") self.world.tick() # You could even add a second tick for good measure if needed # print("Ticking world second time after destruction...") # self.world.tick() print("Post-destruction tick complete.") # --- # Clear Python references AFTER destruction is processed self.cameras = [] self.eval_camera = None self.collision_sensor = None self.lane_invasion_sensor = None self.vehicle = None print("Actor references cleared.") def close(self): self.destroy_sensors() print('carla disconnected.') def terminate(self): self.early_terminate = True def invade(self): self.invaded = True def save_image(self, image): plt.imsave(os.path.join(self.dir, f'ep_{self.episode_number}', f'obs_{self.obs_number:03}.png'), image.transpose(1, 2, 0)) @staticmethod def process_image(image): array = np.frombuffer(image.raw_data, dtype=np.dtype("uint8")) array = np.reshape(array, (image.height, image.width, 4)) array = array[:, :, :3] array = array[:, :, ::-1] array = array.transpose((2, 0, 1)) return array.copy()
and using the following rollout function for ppo agent: ` # rollout for num_steps in the environment def rollout(self, start_global_step):
global_step = start_global_step for step in range(0, self.num_steps): global_step += 1 self.obs[step] = self.next_obs self.commands[step] = self.next_command self.speeds[step] = self.next_speed self.dones[step] = self.next_done # action logic with torch.no_grad(): action, logproba, _, value = self.get_action_and_value( self.obs[step].unsqueeze(0), self.commands[step].unsqueeze(0), self.speeds[step].unsqueeze(0)) self.values[step] = value.view(-1) self.actions[step] = action.view(-1) self.logprobs[step] = logproba.view(-1) # execute the game and log data next_obs, next_command, next_speed, reward, done, info = self.env.step(action.view(-1).cpu().numpy()) self.rewards[step] = torch.tensor(reward).to(self.device) self.next_obs = torch.Tensor(next_obs).to(self.device) self.next_command = torch.Tensor(next_command).to(self.device) self.next_speed = torch.Tensor([next_speed]).to(self.device) self.next_done = torch.tensor(int(done)).to(self.device) if self.next_done.item() == 1: obs, command, speed = self.env.reset() self.next_obs = torch.Tensor(obs).to(self.device) self.next_command = torch.Tensor(command).to(self.device) self.next_speed = torch.Tensor([speed]).to(self.device) if 'distance' in info.keys(): self.writer.add_scalar('charts/distance', info['distance'], global_step) return global_step`
And the third camera is not spawning, I don't know why! I have tried everything! Steps to reproduce Explain the steps needed to reproduce the issue.
Expected behavior A clear and concise description of what you expected to happen.
Logs Paste the server and client logs. Server ones can be found at:
Scripts Add any scripts relevant for the reproducibility of the error.
Screenshots
Additional context Add any other context about the problem here.
The text was updated successfully, but these errors were encountered:
No branches or pull requests
Setup
Describe the setup you are using to run CARLA along with its version:
Describe the bug
I am using the following env.py
from env import routes
and using the following rollout function for ppo agent:
` # rollout for num_steps in the environment
def rollout(self, start_global_step):
And the third camera is not spawning, I don't know why! I have tried everything!
Steps to reproduce
Explain the steps needed to reproduce the issue.
Expected behavior
A clear and concise description of what you expected to happen.
Logs
Paste the server and client logs. Server ones can be found at:
Scripts
Add any scripts relevant for the reproducibility of the error.
Screenshots
Additional context
Add any other context about the problem here.
The text was updated successfully, but these errors were encountered: