ARGS = parser.parse_args()
#### Check the environment's spaces ################################################################
env = gym.make("takeoff-aviary-v0")
print("[INFO] Action space:", env.action_space)
print("[INFO] Observation space:", env.observation_space)
check_env(env, warn=True, skip_render_check=True)
#### Train the model ###############################################################################
if not ARGS.rllib:
model = A2C(MlpPolicy, env, verbose=1)
model.learn(total_timesteps=1000) # e.g. 500000
else:
ray.shutdown()
ray.init(ignore_reinit_error=True)
register_env("takeoff-aviary-v0", lambda _: TakeoffAviary())
config = ppo.DEFAULT_CONFIG.copy()
config["num_workers"] = 2
config["env"] = "takeoff-aviary-v0"
agent = ppo.PPOTrainer(config)
for i in range(10): # e.g. 100
results = agent.train()
print("[INFO] {:d}: episode_reward max {:f} min {:f} mean {:f}".format(i, \
results["episode_reward_max"], results["episode_reward_min"], results["episode_reward_mean"]))
policy = agent.get_policy()
print(policy.model.base_model.summary())
ray.shutdown()
#### Show (and record a video of) the model's performance ##########################################
env = TakeoffAviary(gui=True, record=False)
logger = Logger(logging_freq_hz=int(env.SIM_FREQ / env.AGGR_PHY_STEPS),