from gym.envs.external_games.supertuxkart import SuperTuxKart
import sys
import random
SCREENSHOTS = (len(sys.argv) > 1 and sys.argv[1] == '--screenshots')
FRAMEBUFFER_SIZE = 10
env = SuperTuxKart(spf=0.000887,
framebuffer_size=FRAMEBUFFER_SIZE,
width=800,
height=600,
level='scotland',
mode='time',
speedup=1,
observe_performance=False)
ACTIONS_WORDS = [
'none', 'throttle', 'brake', 'use_powerup', 'drift', 'nitro', 'left',
'right'
]
ACTIONS_NUMBERS = list(range(len(ACTIONS_WORDS)))
screenshot_id = 0
def make_action(action):
global screenshot_id
print(action)
ob, reward, done, scores = env.step(action)
print(scores)
print(env.fps)
import time
from gym import logger
from gym.envs.external_games.supertuxkart import SuperTuxKart
FRAMEBUFFER_SIZE = 4
env = SuperTuxKart(spf=0.000887,
framebuffer_size=FRAMEBUFFER_SIZE,
width=800,
height=600,
level='scotland',
mode='time')
logger.set_level(logger.DEBUG)
last_gt = None
last_st = None
delta_gts = []
delta_sts = []
step_times = []
tot_game_fps = []
tot_gym_fps = []
for i in range(100):
before = time.time()
ob, reward, done, scores = env.step('throttle')
after = time.time()
gt = scores['game_time']
st = scores['system_time']
if last_st is not None and last_gt is not None:
delta_gt = gt - last_gt
next_state_values = next_state_values.detach() # return the tensor without connection to its calculation history
expected_state_action_values = next_state_values * GAMMA + rewards_v
return nn.MSELoss()(state_action_values, expected_state_action_values)
# **********************************************************************************************************************
# * TRAINING STARTS *
# **********************************************************************************************************************
if __name__ == "__main__":
# set the device -> GPU or CPU
device = "cuda"
# create the wrapped environment
env = SuperTuxKart(spf=0.000887, framebuffer_size=FRAME_BUFFER_SIZE, width=SCREEN_WIDTH, height=SCREEN_HEIGHT,
level=GAME_LEVEL, mode=GAME_MODE, speedup=GAME_SPEED, observe_performance=OBS_PERFORMANCE,
performance_window_size=WINDOW_ACTIONS, performance_window_overlap=OVERLAP_ACTIONS, laps=LAPS)
# **********************************************************************************************************************
# * DEFINE THRESHOLD *
# **********************************************************************************************************************
if NUM_PREGAMES > 0:
print('******************************* STARTING PRE GAMES TO SET THE THRESHOLD *******************************')
# create the net and the target net
net = dqn_model.DQN((FRAME_BUFFER_SIZE, SCREEN_WIDTH, SCREEN_HEIGHT), len(POSSIBLE_ACTIONS)).to(device)
tgt_net = dqn_model.DQN((FRAME_BUFFER_SIZE, SCREEN_WIDTH, SCREEN_HEIGHT), len(POSSIBLE_ACTIONS)).to(device)
net.load_state_dict(torch.load(DEFAULT_ENV_NAME + "-best_RL-baseline.dat"))
tgt_net.load_state_dict(torch.load(DEFAULT_ENV_NAME + "-best_RL-baseline.dat"))
print(net)
return ['throttle', 'left']
elif centering < -self.CENTERING_TOLERANCE:
return ['throttle', 'right']
else:
return 'throttle'
if __name__ == '__main__':
# You can set the level to logger.DEBUG or logger.WARN if you
# want to change the amount of output.
logger.set_level(logger.DEBUG)
env = SuperTuxKart(spf=0.000887,
framebuffer_size=4,
width=800,
height=600,
level='scotland',
mode='time',
speedup=1,
observe_performance=False)
# You provide the directory to write to (can be an existing
# directory, including one with existing data -- all monitor files
# will be namespaced). You can also dump to a tempdir if you'd
# like: tempfile.mkdtemp().
outdir = '/tmp/random-agent-results'
agent = SuperTuxKartAgent(env.action_space)
episode_count = 5
reward = 0
scores = None
[ACTIONS[5]], # RIGHT
[ACTIONS[1], ACTIONS[4]], # UP + LEFT
[ACTIONS[1], ACTIONS[5]], # UP + RIGHT
[ACTIONS[1], ACTIONS[2]], # UP + POWER
[ACTIONS[1], ACTIONS[3]] # UP + NITRO
]
# set the device -> GPU or CPU
device = "cuda"
# create the wrapped environment
env = SuperTuxKart(spf=0.000887,
framebuffer_size=FRAME_BUFFER_SIZE,
width=SCREEN_WIDTH,
height=SCREEN_HEIGHT,
level=GAME_LEVEL,
mode=GAME_MODE,
speedup=GAME_SPEED,
observe_performance=True,
performance_window_size=2,
performance_window_overlap=1,
laps=1)
all_games_positions = []
all_consumptions = []
for game in range(1000):
env.reset()
ob, r, is_done, info = env.step(ACTIONS[0])
while info['game_time'] < 0:
ob, r, is_done, info = env.step(ACTIONS[0])
starting_time = info['game_time']