def render(self, mode='human'):
assert mode in ['human', 'human-vsync', 'rgb_array', 'manual']
if mode == 'human-vsync':
self.vsync = True
if mode == 'manual':
self.manual = 1
if self.viewer is None:
self.viewer = Viewer(window_h_size, window_v_size, self.manual,
self.vsync)
[self.track_label, self.time_label, self.episode_label] = \
self.viewer.labels(self.main_batch, self.label_font_size, self.track_name, self.t, self.episode)
# redefine draw event
pyglet.gl.glClearColor(1, 1, 1, 1) # white background
self.viewer.toDraw = [self.track_batch, self.main_batch]
if debug:
self.viewer.toDraw.append(self.debug_batch)
if self.manual:
self.keys = pyglet.window.key.KeyStateHandler()
if self.viewer.is_open:
if mode == 'human' or mode == 'manual':
self.viewer.render()
elif mode == 'rgb_array':
rgb = self.viewer.render(True)
return rgb
if mode == 'manual':
self.viewer.window.push_handlers(self.keys)
def render(self, mode='human'):
'''
'human' - runs as fast as possible
'human-vsync' - runs at screen framerate (timestep per frame is still 1/60) also includes anti aliasing
'rgb_array' - returns an RGB array instead of rendering the output in the window
'manual' - overrides input actions with keyboard inputs, renders like human-vsync
'''
assert mode in ['human', 'human-vsync', 'rgb_array', 'manual']
if mode == 'human-vsync':
self.vsync = True
if mode == 'manual':
self.manual = 1
if self.viewer is None:
self.viewer = Viewer(window_h_size, window_v_size, self.manual,
self.vsync)
[self.track_label, self.time_label, self.episode_label] = \
self.viewer.labels(self.main_batch, 36, self.track_name, self.t, self.episode)
# redefine draw event
pyglet.gl.glClearColor(1, 1, 1, 1) # white background
self.viewer.toDraw = [self.track_batch, self.main_batch]
if debug:
self.viewer.toDraw.append(self.debug_batch)
if self.manual:
self.keys = pyglet.window.key.KeyStateHandler()
if self.viewer.is_open:
if mode == 'human' or mode == 'manual':
self.viewer.render()
elif mode == 'rgb_array':
rgb = self.viewer.render(True)
return rgb
if mode == 'manual':
self.viewer.window.push_handlers(self.keys)
class CarAIEnv(gym.Env):
metadata = {
'render.modes': ['human', 'human-vsync', 'rgb_array', 'manual'],
'video.frames_per_second': 60
}
'''
human - displays the environment, no vsync, timestep as fast as possible,
human-vsync - displays the environment, vsync
rgb_array - Outputs rgb array instead of rendering
manual - human-vsync but actions are taken from arrow keys and not from the inputs. Testing mode.
'''
def __init__(self):
pyglet.resource.path = ['gym_carai/envs/resources']
pyglet.resource.reindex()
self.mode = 'not_simple'
self.label_font_size = 36
self.viewer = None
self.Terminate = False
self.manual = None
self.keys = None
self.vsync = False
# 3 batches, one for car, one for obstacles, one for debug features
self.main_batch = pyglet.graphics.Batch()
self.track_batch = pyglet.graphics.Batch()
self.debug_batch = pyglet.graphics.Batch()
# Labels for score and level, scaled based upon score label font size
self.score = 0
self.t = 0
self.episode = 1
self.done = 0
self.reward = 0
self.JStar = 0
self.track_label = None
self.time_label = None
self.episode_label = None
self.action_space = spaces.Box(np.array([-1]), np.array(
[+1])) # steering only, -1 to +1 on one action
self.track_name = 'roundSimpleTrackBC'
# define functions
self.walls, self.checkpoints, car_position = generate_track(
'gym_carai/envs/resources/' + self.track_name + '.csv',
self.track_batch)
self.current_checkpoint = 1
self.car_obj = Car(car_position,
debug_batch=self.debug_batch,
main_batch=self.main_batch,
mode=self.mode)
self.car_bumpers = [
self.car_obj.Bumper, self.car_obj.SideL, self.car_obj.SideR,
self.car_obj.Rear
]
# all objects requiring .update()
self.envObjects = [self.car_obj]
self.sensors = self.car_obj.sensors
self.observation_space = spaces.Box(
np.zeros(len(self.sensors)),
self.car_obj.sensorRange * np.ones(len(self.sensors)))
self.observations = np.array([np.zeros(len(self.sensors))])
def step(self, action, dt):
""""observation (object): agent's observation of the current environment
reward (float) : amount of reward returned after previous action
done (bool): whether the episode has ended, in which case further step() calls will return undefined results
info (dict): contains auxiliary diagnostic information (helpful for debugging, and sometimes learning)"""
self.reward = -0.05 # reward for current state if not at a checkpoint
if self.manual:
if self.keys[pyglet.window.key.LEFT] == 1:
if self.keys[pyglet.window.key.RIGHT] == 1:
action[0] = 0
else:
action[0] = -1
elif self.keys[pyglet.window.key.RIGHT] == 1:
action[0] = 1
else:
action[0] = 0
if self.keys[pyglet.window.key.UP] == 1:
if self.keys[pyglet.window.key.DOWN] == 1:
action[1] = 0
else:
action[1] = 1
elif self.keys[pyglet.window.key.DOWN] == 1:
action[1] = -1
else:
action[1] = 0
# if self.keys[pyglet.window.key.SPACE] == 1:
# action[2] = 1
# else:
# action[2] = 0
for obj in self.envObjects:
obj.update(dt, action)
for obj in self.walls:
for car_bumper in self.car_bumpers:
tf = line_overlapping(car_bumper.line(), obj.line())
if tf:
self.done = 1
for obj in self.checkpoints:
if obj.id == self.current_checkpoint:
for car_bumper in self.car_bumpers:
tf = line_overlapping(car_bumper.line(), obj.line())
if tf:
self.current_checkpoint += 1
if self.current_checkpoint > len(self.checkpoints):
self.current_checkpoint -= len(self.checkpoints)
self.reward = 0
current_sensor_number = 0
for sensor in self.sensors:
min_distance = sensor.sensor_range
closest_col_loc = [-50, -50]
for obj in self.walls:
tf, dist, col_loc = line_overlapping(sensor.line(),
obj.line(),
get_dist=True,
printT=True)
if dist is not None and tf is True:
if dist < min_distance:
min_distance = dist
closest_col_loc = col_loc
sensor.collision_marker.x = closest_col_loc[0]
sensor.collision_marker.y = closest_col_loc[1]
self.observations[0][current_sensor_number] = min_distance
current_sensor_number += 1
self.t += dt
if self.time_label:
self.time_label.text = "Current Episode Time: " + str(round(
self.t))
if self.viewer:
if not self.Terminate:
self.Terminate = self.viewer.Terminate
if self.done:
self.reward = -50 # - 100/self.t # penalty for hitting wall, higher penalty if wall is hit early
self.JStar = 0 # All rewards are negative so JStar is zero.
return self.observations, self.reward, self.done, {
't': self.t,
'JStar': self.JStar
}, self.Terminate
def reset(self):
self.score = 0
self.current_checkpoint = 1
self.t = 0
self.reward = 0
self.done = 0
for obj in self.envObjects:
obj.reset()
if self.viewer:
self.episode_label.text = "Current episode:" + str(self.episode)
self.episode += 1
def render(self, mode='human'):
assert mode in ['human', 'human-vsync', 'rgb_array', 'manual']
if mode == 'human-vsync':
self.vsync = True
if mode == 'manual':
self.manual = 1
if self.viewer is None:
self.viewer = Viewer(window_h_size, window_v_size, self.manual,
self.vsync)
[self.track_label, self.time_label, self.episode_label] = \
self.viewer.labels(self.main_batch, self.label_font_size, self.track_name, self.t, self.episode)
# redefine draw event
pyglet.gl.glClearColor(1, 1, 1, 1) # white background
self.viewer.toDraw = [self.track_batch, self.main_batch]
if debug:
self.viewer.toDraw.append(self.debug_batch)
if self.manual:
self.keys = pyglet.window.key.KeyStateHandler()
if self.viewer.is_open:
if mode == 'human' or mode == 'manual':
self.viewer.render()
elif mode == 'rgb_array':
rgb = self.viewer.render(True)
return rgb
if mode == 'manual':
self.viewer.window.push_handlers(self.keys)
def close(self):
if self.viewer:
self.viewer.close()