def test_render_heading_control(self):
self.setUp(plane=aircraft.a320, task_type=tasks.TurnHeadingControlTask,
shaping=tasks.Shaping.EXTRA_SEQUENTIAL)
agent = RandomAgent(self.env.action_space)
render_every = 5
report_every = 20
EPISODES = 50
for _ in range(EPISODES):
ep_reward = 0
done = False
state = self.env.reset()
self.env.render(mode='flightgear')
step_number = 0
while not done:
action = agent.act(state)
state, reward, done, info = self.env.step(action)
ep_reward += reward
if step_number % render_every == 0:
self.env.render(mode='flightgear')
if step_number % report_every == 0:
heading_target = tasks.HeadingControlTask.target_track_deg
print(f'time:\t{self.env.sim.get_sim_time()} s')
print(f'last reward:\t{reward}')
print(f'episode reward:\t{ep_reward}')
print(f'gear status:\t{self.env.sim[prp.gear]}')
print(f'thrust eng0:\t{self.env.sim[prp.engine_thrust_lbs]}')
print(f'thrust eng1:\t {self.env.sim[prp.Property("propulsion/engine[1]/thrust-lbs", "")]}')
print(f'heading:\t{self.env.sim[prp.heading_deg]}')
print(f'target heading:\t{self.env.sim[heading_target]}')
print('\n')
step_number += 1
print(f'***\n'
f'EPISODE REWARD: {ep_reward}\n'
f'***\n')
def test_render_steady_level_flight_random(self):
""" Runs steady level flight task with a random agent. """
self.setUp(task_type=tasks.HeadingControlTask)
agent = RandomAgent(self.env.action_space)
render_every = 5
ep_reward = 0
done = False
state = self.env.reset()
step_number = 0
while not done:
action = agent.act(state)
state, reward, done, info = self.env.step(action)
ep_reward += reward
if step_number % render_every == 0:
self.env.render(mode='human')
step_number += 1
def init_and_reset_env(self, env: JsbSimEnv):
self.assertIsInstance(env.task, HeadingControlTask)
# we interact at 5 Hz, so we expect the sim to run 12 timesteps per
# interaction since it runs at 120 Hz
self.assertEqual(12, env.sim_steps_per_agent_step)
# we init a random agent with a seed
agent = RandomAgent(action_space=env.action_space)
self.assertEqual(env.action_space, agent.action_space)
# this task has an action space of three controls: aileron, elevator, rudder
expected_num_actions = 3
self.assertEqual(expected_num_actions, len(agent.action_space.low))
# we see that the action space has the correct low and high range of +-1.0
expect_low = np.array([-1.0] * expected_num_actions)
expect_high = np.array([1.0] * expected_num_actions)
np.testing.assert_array_almost_equal(expect_high,
env.action_space.high)
np.testing.assert_array_almost_equal(expect_low, env.action_space.low)
# we reset the env and receive the first state; the env is now ready
state = env.reset()
self.assertEqual(len(env.observation_space.low), len(state))
# we close the env and JSBSim closes with it
env.close()
self.assertIsNone(env.sim.jsbsim)
class TestRandomAgent(unittest.TestCase):
def setUp(self):
self.action_space = FlightTaskStub().get_action_space()
self.agent = RandomAgent(action_space=self.action_space)
def test_act_generates_valid_actions(self):
num_test_actions = 5
for _ in range(num_test_actions):
action = self.agent.act(None)
self.assertTrue(self.action_space.contains(action))
def test_run_episode_steady_level_flight_no_render(self):
self.setUp(task_type=tasks.HeadingControlTask)
agent = RandomAgent(self.env.action_space)
report_every = 20
EPISODES = 10
for _ in range(EPISODES):
ep_reward = 0
done = False
state = self.env.reset()
step_number = 0
while not done:
action = agent.act(state)
state, reward, done, info = self.env.step(action)
ep_reward += reward
if step_number % report_every == 0:
print(f'time:\t{self.env.sim.get_sim_time()} s')
print(f'last reward:\t{reward}')
print(f'episode reward:\t{ep_reward}')
step_number += 1
def take_step_with_random_agent(self, env: JsbSimEnv):
agent = RandomAgent(action_space=env.action_space)
# we set up for a loop through one episode
first_state = env.reset()
# we take a single step
action = agent.act(first_state)
state, reward, done, info = env.step(action)
# we see the state has changed
self.assertEqual(first_state.shape, state.shape)
self.assertTrue(np.any(np.not_equal(first_state, state)),
msg='state should have changed after simulation step')
expected_time_step_size = env.sim_steps_per_agent_step / env.JSBSIM_DT_HZ
self.assertAlmostEqual(expected_time_step_size, env.sim.get_sim_time())
self.assertFalse(done,
msg='episode is terminal after only a single step')
# the aircraft engines are running, as per initial conditions
self.assertNotAlmostEqual(env.sim[prp.engine_thrust_lbs], 0)
env.close()
def setUp(self):
self.action_space = FlightTaskStub().get_action_space()
self.agent = RandomAgent(action_space=self.action_space)